app.py

"""
BrainConnect-ASD — Scanner-site-invariant ASD detection from fMRI.
"""
from __future__ import annotations

import io
import os
from pathlib import Path

import numpy as np
import torch
import gradio as gr

from _charts import VAL_B64, AUC_B64

_WINDOW_LEN   = 50
_STEP         = 3
_MAX_WINDOWS  = 30
_FC_THRESHOLD = 0.2

# ── Atlas configurations ────────────────────────────────────────────────────
# CC200 → Yeo 7-network parcellation (approximate ROI ordering)
_ATLAS_CFG = {
    "cc200": {
        "n_rois":      200,
        "label":       "CC200",
        "net_names":   ["DMN", "Salience", "Frontoparietal", "Sensorimotor", "Visual", "Dorsal Attn", "Subcortical"],
        "net_bounds":  [0, 38, 69, 99, 137, 165, 180, 200],
        "net_colors":  ["#e63946", "#f4a261", "#457b9d", "#2dc653", "#a8dadc", "#8b5cf6", "#6b7280"],
        "ckpts": {
            "CALTECH":  Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_caltech/brain-gcn-epoch=020-val_auc=0.953.ckpt"),
            "CMU":      Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_cmu/brain-gcn-epoch=001-val_auc=0.893.ckpt"),
            "KKI":      Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_kki/brain-gcn-epoch=014-val_auc=0.917.ckpt"),
            "LEUVEN_1": Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_leuven_1/brain-gcn-epoch=004-val_auc=0.917.ckpt"),
            "LEUVEN_2": Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_leuven_2/brain-gcn-epoch=005-val_auc=0.888.ckpt"),
            "MAX_MUN":  Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_max_mun/brain-gcn-epoch=005-val_auc=0.858.ckpt"),
            "NYU":      Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_nyu/brain-gcn-epoch=067-val_auc=0.964.ckpt"),
            "OHSU":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_ohsu/brain-gcn-epoch=004-val_auc=0.858.ckpt"),
            "OLIN":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_olin/brain-gcn-epoch=003-val_auc=0.970.ckpt"),
            "PITT":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_pitt/brain-gcn-epoch=009-val_auc=0.935.ckpt"),
            "SBL":      Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_sbl/brain-gcn-epoch=021-val_auc=0.876.ckpt"),
            "SDSU":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_sdsu/brain-gcn-epoch=001-val_auc=0.864.ckpt"),
            "STANFORD": Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_stanford/brain-gcn-epoch=002-val_auc=0.923.ckpt"),
            "TRINITY":  Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_trinity/brain-gcn-epoch=006-val_auc=0.888.ckpt"),
            "UCLA_1":   Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_ucla_1/brain-gcn-epoch=054-val_auc=0.976.ckpt"),
            "UCLA_2":   Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_ucla_2/brain-gcn-epoch=055-val_auc=0.863.ckpt"),
            "UM_1":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_um_1/brain-gcn-epoch=013-val_auc=0.959.ckpt"),
            "UM_2":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_um_2/brain-gcn-epoch=005-val_auc=0.899.ckpt"),
            "USM":      Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_usm/brain-gcn-epoch=020-val_auc=0.970.ckpt"),
            "YALE":     Path("checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_yale/brain-gcn-epoch=055-val_auc=0.964.ckpt"),
        },
    },
    "aal": {
        "n_rois":      116,
        "label":       "AAL-116",
        # Approximate Yeo-7 parcellation for AAL-116 anatomical ordering:
        # Frontal/FPN (1-28), Sensorimotor (29-40), DMN parietal (41-60),
        # Temporal/DMN (61-76), Subcortical (77-90), Occipital/Visual (91-116)
        "net_names":   ["Frontoparietal", "Sensorimotor", "Dorsal Attn", "DMN", "Salience", "Subcortical", "Visual"],
        "net_bounds":  [0, 20, 34, 50, 68, 80, 92, 116],
        "net_colors":  ["#457b9d", "#2dc653", "#8b5cf6", "#e63946", "#f4a261", "#6b7280", "#a8dadc"],
        "ckpts": {
            "CALTECH":  Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_caltech/brain-gcn-epoch=003-val_auc=0.822.ckpt"),
            "CMU":      Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_cmu/brain-gcn-epoch=004-val_auc=0.775.ckpt"),
            "KKI":      Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_kki/brain-gcn-epoch=022-val_auc=0.834.ckpt"),
            "LEUVEN_1": Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_leuven_1/brain-gcn-epoch=001-val_auc=0.858.ckpt"),
            "LEUVEN_2": Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_leuven_2/brain-gcn-epoch=007-val_auc=0.846.ckpt"),
            "MAX_MUN":  Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_max_mun/brain-gcn-epoch=056-val_auc=0.769.ckpt"),
            "NYU":      Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_nyu/brain-gcn-epoch=011-val_auc=0.740.ckpt"),
            "OHSU":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_ohsu/brain-gcn-epoch=006-val_auc=0.799.ckpt"),
            "OLIN":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_olin/brain-gcn-epoch=008-val_auc=0.846.ckpt"),
            "PITT":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_pitt/brain-gcn-epoch=001-val_auc=0.888.ckpt"),
            "SBL":      Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_sbl/brain-gcn-epoch=018-val_auc=0.828.ckpt"),
            "SDSU":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_sdsu/brain-gcn-epoch=005-val_auc=0.746.ckpt"),
            "STANFORD": Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_stanford/brain-gcn-epoch=002-val_auc=0.852.ckpt"),
            "TRINITY":  Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_trinity/brain-gcn-epoch=001-val_auc=0.834.ckpt"),
            "UCLA_1":   Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_ucla_1/brain-gcn-epoch=000-val_auc=0.846.ckpt"),
            "UCLA_2":   Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_ucla_2/brain-gcn-epoch=000-val_auc=0.813.ckpt"),
            "UM_1":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_um_1/brain-gcn-epoch=051-val_auc=0.828.ckpt"),
            "UM_2":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_um_2/brain-gcn-epoch=001-val_auc=0.822.ckpt"),
            "USM":      Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_usm/brain-gcn-epoch=006-val_auc=0.805.ckpt"),
            "YALE":     Path("checkpoints/aal/adv_brain_mode_k32_site_aal_loso_yale/brain-gcn-epoch=054-val_auc=0.870.ckpt"),
        },
    },
    "ho": {
        "n_rois":      111,
        "label":       "Harvard-Oxford",
        "net_names":   ["Frontoparietal", "Sensorimotor", "DMN", "Salience", "Subcortical", "Visual", "Temporal"],
        "net_bounds":  [0, 18, 30, 48, 68, 80, 96, 111],
        "net_colors":  ["#457b9d", "#2dc653", "#e63946", "#f4a261", "#6b7280", "#a8dadc", "#8b5cf6"],
        "ckpts": {
            "CALTECH":  Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_caltech/brain-gcn-epoch=013-val_auc=0.888.ckpt"),
            "CMU":      Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_cmu/brain-gcn-epoch=011-val_auc=0.852.ckpt"),
            "KKI":      Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_kki/brain-gcn-epoch=059-val_auc=0.917.ckpt"),
            "LEUVEN_1": Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_leuven_1/brain-gcn-epoch=021-val_auc=0.899.ckpt"),
            "LEUVEN_2": Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_leuven_2/brain-gcn-epoch=055-val_auc=0.905.ckpt"),
            "MAX_MUN":  Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_max_mun/brain-gcn-epoch=003-val_auc=0.882.ckpt"),
            "NYU":      Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_nyu/brain-gcn-epoch=017-val_auc=0.882.ckpt"),
            "OHSU":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_ohsu/brain-gcn-epoch=010-val_auc=0.882.ckpt"),
            "OLIN":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_olin/brain-gcn-epoch=024-val_auc=0.929.ckpt"),
            "PITT":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_pitt/brain-gcn-epoch=018-val_auc=0.882.ckpt"),
            "SBL":      Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_sbl/brain-gcn-epoch=003-val_auc=0.893.ckpt"),
            "SDSU":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_sdsu/brain-gcn-epoch=095-val_auc=0.935.ckpt"),
            "STANFORD": Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_stanford/brain-gcn-epoch=002-val_auc=0.888.ckpt"),
            "TRINITY":  Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_trinity/brain-gcn-epoch=021-val_auc=0.864.ckpt"),
            "UCLA_1":   Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_ucla_1/brain-gcn-epoch=009-val_auc=0.817.ckpt"),
            "UCLA_2":   Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_ucla_2/brain-gcn-epoch=001-val_auc=0.797.ckpt"),
            "UM_1":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_um_1/brain-gcn-epoch=005-val_auc=0.852.ckpt"),
            "UM_2":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_um_2/brain-gcn-epoch=006-val_auc=0.870.ckpt"),
            "USM":      Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_usm/brain-gcn-epoch=000-val_auc=0.840.ckpt"),
            "YALE":     Path("checkpoints/ho/adv_brain_mode_k32_site_ho_loso_yale/brain-gcn-epoch=004-val_auc=0.876.ckpt"),
        },
    },
}

# Resolve active atlas config by ROI count
_ROI_TO_ATLAS = {cfg["n_rois"]: key for key, cfg in _ATLAS_CFG.items()}

# Legacy aliases kept for backward compat
_NET_NAMES  = _ATLAS_CFG["cc200"]["net_names"]
_NET_BOUNDS = _ATLAS_CFG["cc200"]["net_bounds"]
_NET_COLORS = _ATLAS_CFG["cc200"]["net_colors"]
_CKPTS      = _ATLAS_CFG["cc200"]["ckpts"]

# ── preprocessing ──────────────────────────────────────────────────────────

def _zscore(bold):
    mean = bold.mean(0, keepdims=True)
    std  = bold.std(0, keepdims=True)
    std[std < 1e-8] = 1.0
    return ((bold - mean) / std).astype(np.float32)

def _fc(bold):
    fc = np.corrcoef(bold.T).astype(np.float32)
    np.nan_to_num(fc, copy=False)
    return fc

def _windows(bold):
    T, N = bold.shape
    starts = list(range(0, T - _WINDOW_LEN + 1, _STEP))
    w = np.stack([bold[s:s+_WINDOW_LEN].std(0) for s in starts]).astype(np.float32)
    if len(w) >= _MAX_WINDOWS:
        return w[:_MAX_WINDOWS]
    return np.concatenate([w, np.repeat(w[-1:], _MAX_WINDOWS - len(w), 0)])

def preprocess(bold):
    bold = _zscore(bold)
    fc   = _fc(bold)
    fc   = np.arctanh(np.clip(fc, -0.9999, 0.9999))
    adj  = np.where(np.abs(fc) >= _FC_THRESHOLD, fc, 0.0).astype(np.float32)
    bw   = _windows(bold)
    return torch.FloatTensor(bw).unsqueeze(0), torch.FloatTensor(adj).unsqueeze(0)

# ── LLM (Qwen2.5-7B fine-tuned on AMD MI300X, served via vLLM on MI300X) ───

_VLLM_URL   = os.environ.get("VLLM_BASE_URL", "")
_LLM_MODEL  = "lablab-ai-amd-developer-hackathon/asd-interpreter-merged"
_HF_TOKEN   = os.environ.get("HF_TOKEN", "")

# Pre-generated reports for demo subjects (instant display, no LLM latency)
_DEMO_LLM_CACHE = {
    "sample_asd_stanford.1D": """ICD-10: F84.0 (Childhood Autism) / F84.1 (Atypical Autism)
Ensemble Confidence: HIGH · p(ASD) = 0.841 · 19/20 site-blind models agree

IMPRESSION
Strong ASD-consistent functional connectivity profile. The ensemble shows high cross-site agreement, indicating the pattern is robust to scanner and acquisition differences across the 20 ABIDE sites.

CONNECTIVITY FINDINGS
• Default Mode Network shows reduced long-range coherence, consistent with atypical self-referential processing reported in ASD
• Elevated saliency in Frontoparietal ↔ Subcortical pathways, suggesting atypical executive-limbic coupling
• Visual network exhibits disproportionate connectivity weight relative to DMN — consistent with sensory hypersensitivity profiles in ASD

CROSS-SITE CONSISTENCY
19/20 site-blind models agree — pattern is not attributable to scanner artifacts (Stanford site held out during training).

SUPPORTING LITERATURE
• Rudie et al. 2012 — Reduced functional integration in ASD
• Washington et al. 2014 — Dysmaturation of the default mode network in autism

AI-assisted screening only · Not a clinical diagnosis · Requires full ADOS-2 and developmental history evaluation""",

    "sample_tc_yale.1D": """ICD-10: Z03.89 (No diagnosis) — Typical Connectivity Profile
Ensemble Confidence: HIGH (TC) · p(ASD) = 0.143 · 18/20 site-blind models predict Typical Control

IMPRESSION
Connectivity profile is consistent with neurotypical development. The ensemble shows strong agreement against ASD classification across held-out sites.

CONNECTIVITY FINDINGS
• Default Mode Network coherence within expected range for age-matched neurotypical controls
• Frontoparietal ↔ DMN anticorrelation preserved — consistent with intact task-positive/task-negative network segregation
• Salience network lateralization within normative bounds

CROSS-SITE CONSISTENCY
18/20 site-blind models predict Typical Control — Yale site held out during training, result generalizes across scanner environments.

AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with full clinical assessment""",

    "sample_borderline_trinity.1D": """ICD-10: F84.5 (Asperger Syndrome) — Borderline / Uncertain
Ensemble Confidence: LOW/UNCERTAIN · p(ASD) = 0.523 · 11/20 site-blind models predict ASD

IMPRESSION
Borderline connectivity profile with high inter-model variance. The ensemble is split, indicating this subject falls near the decision boundary. Clinical evaluation is essential — GCN classification alone is insufficient for borderline cases.

CONNECTIVITY FINDINGS
• Default Mode Network shows mild coherence reduction, below the threshold seen in clear ASD cases
• Frontoparietal network saliency is elevated but inconsistent across site-blind models
• Salience network shows atypical lateralization in a subset of models only

CROSS-SITE CONSISTENCY
11/20 models predict ASD, 9/20 predict Typical Control. High variance suggests scanner-site sensitivity — Trinity site held out during training.

RECOMMENDATION
Full neuropsychological evaluation recommended including ADOS-2, ADI-R, and cognitive assessment. Borderline fMRI profiles are common in high-functioning ASD and require multi-modal diagnostic workup.

AI-assisted screening only · Not a clinical diagnosis"""
}

_SYSTEM_PROMPT = (
    "You are a clinical AI assistant specializing in functional MRI brain "
    "connectivity analysis for autism spectrum disorder (ASD) diagnosis support. "
    "You interpret outputs from a validated graph neural network (GCN) trained on "
    "the ABIDE I dataset (1,102 subjects, 20 acquisition sites) and provide structured "
    "clinical summaries for neurologists and psychiatrists. "
    "CRITICAL RULES: (1) Only reference brain networks, connectivity patterns, and "
    "statistics that are explicitly provided in the input report — do NOT invent or "
    "hallucinate network names, connectivity findings, or numeric values. "
    "(2) Base every observation directly on the per-network saliency scores and "
    "ensemble probabilities given in the input. (3) If a network is not listed in the "
    "input, do not mention it. (4) Always clarify findings are AI-assisted and require "
    "full clinical assessment. You do not make a diagnosis."
)

def _llm_report(p_mean: float, per_model: list, net_saliency: dict | None = None) -> str:
    consensus = sum(1 for _, p in per_model if p > 0.5)
    per_model_str = "\n".join(
        f"  {s}-blind: {'ASD' if v > 0.5 else 'TC'} (p={v:.3f})" for s, v in per_model
    )
    conf_label = (
        "HIGH" if p_mean >= 0.75 else
        "MODERATE" if p_mean >= 0.6 else
        "LOW / UNCERTAIN" if p_mean >= 0.4 else
        "MODERATE (TC)" if p_mean >= 0.25 else "HIGH (TC)"
    )

    sal_section = ""
    if net_saliency:
        sorted_nets = sorted(net_saliency.items(), key=lambda x: x[1], reverse=True)
        sal_lines = "\n".join(
            f"  {name}: {score:.5f}" for name, score in sorted_nets
        )
        sal_section = (
            f"\nPer-Network Gradient Saliency (ranked high→low, actual GCN values):\n"
            f"{sal_lines}\n"
            f"[ONLY reference these networks with these exact values — no others.]\n"
        )

    user_msg = (
        f"Brain Connectivity GCN Analysis Report\n{'='*40}\n"
        f"Dataset           : ABIDE I · 1,102 subjects · 20 acquisition sites\n"
        f"p(ASD)            : {p_mean:.3f}\n"
        f"Confidence Level  : {conf_label}\n"
        f"Model Consensus   : {consensus}/{len(per_model)} site-blind models predict ASD\n"
        f"{sal_section}\n"
        f"Per-Model Breakdown (LOSO ensemble):\n{per_model_str}\n\n"
        f"Provide a structured clinical interpretation referencing ONLY the networks "
        f"and values listed above. Do not mention any network not in this report."
    )
    try:
        from openai import OpenAI
        if _VLLM_URL:
            # Live AMD MI300X inference via vLLM
            client = OpenAI(base_url=_VLLM_URL, api_key="not-required", timeout=5.0)
            model_id = _LLM_MODEL
        else:
            # Fallback: HF Inference API
            from huggingface_hub import InferenceClient as _HFClient
            client = _HFClient(model=_LLM_MODEL, token=_HF_TOKEN or None)
            response = client.chat_completion(
                messages=[
                    {"role": "system", "content": _SYSTEM_PROMPT},
                    {"role": "user",   "content": user_msg},
                ],
                max_tokens=512, temperature=0.1,
            )
            return response.choices[0].message.content.strip()
        messages = [
            {"role": "system", "content": _SYSTEM_PROMPT},
            {"role": "user",   "content": user_msg},
        ]
        response = client.chat.completions.create(
            model=model_id, messages=messages, max_tokens=512, temperature=0.1
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        # Fallback to cached reports for known demo subjects
        import os as _os
        return "[LLM unavailable — AMD MI300X endpoint offline. Please try again shortly.]"

# ── model loading ──────────────────────────────────────────────────────────

_model_cache: dict[str, list] = {}
_result_cache: dict[str, tuple[str, str, str, object]] = {}

def get_models(atlas: str = "cc200"):
    global _model_cache
    if atlas in _model_cache:
        return _model_cache[atlas]
    from brain_gcn.tasks import ClassificationTask
    cfg = _ATLAS_CFG.get(atlas, _ATLAS_CFG["cc200"])
    models = []
    for site, ckpt in cfg["ckpts"].items():
        if not ckpt.exists():
            continue
        task = ClassificationTask.load_from_checkpoint(str(ckpt), map_location="cpu", strict=False)
        task.eval()
        models.append((site, task))
    _model_cache[atlas] = models
    return models

# ── gradient saliency ──────────────────────────────────────────────────────

def _compute_saliency(bw_t, adj_t, models):
    # Cap at 2 models — backward pass on CPU is slow
    sample = models[:2] if len(models) > 2 else models
    maps = []
    for _, task in sample:
        try:
            adj = adj_t.clone().detach().requires_grad_(True)
            bw  = bw_t.clone().detach()
            with torch.enable_grad():
                out = task.model(bw, adj)
                logits = out[0] if isinstance(out, tuple) else out
                prob = torch.softmax(logits, -1)[0, 1]
                prob.backward()
            if adj.grad is not None:
                maps.append(adj.grad[0].abs().detach().cpu().numpy())
        except Exception as e:
            print(f"[saliency model] {e}")
            continue
    if not maps:
        n = adj_t.shape[-1]
        return np.zeros((n, n), dtype=np.float32)
    sal = np.mean(maps, axis=0)
    return (sal + sal.T) / 2

# Approximate MNI centroids for each CC200 network (mm), used for 3D brain view
_NET_MNI = np.array([
    [ -1, -52,  28],   # DMN        (PCC)
    [  2,  18,  30],   # Salience   (dACC)
    [ 44,  36,  28],   # Frontoparietal (DLPFC)
    [  0, -18,  62],   # Sensorimotor  (SMA/M1)
    [  0, -82,   8],   # Visual     (occipital)
    [ 28, -58,  50],   # Dorsal Attn (IPS)
    [ 14,   4,   4],   # Subcortical (thalamus)
], dtype=np.float32)

def _saliency_figure(sal, p_mean, net_names=None, net_bounds=None, net_colors=None):
    import matplotlib
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt
    from mpl_toolkits.mplot3d import Axes3D  # noqa: F401
    from mpl_toolkits.mplot3d.art3d import Line3DCollection
    from PIL import Image

    _nn = net_names  if net_names  is not None else _NET_NAMES
    _nb = net_bounds if net_bounds is not None else _NET_BOUNDS
    _nc = net_colors if net_colors is not None else _NET_COLORS
    n_nets = len(_nn)

    # Aggregate N×N saliency → 7×7 network-level matrix
    net_sal = np.zeros((n_nets, n_nets))
    for i, (s1, e1) in enumerate(zip(_nb[:-1], _nb[1:])):
        for j, (s2, e2) in enumerate(zip(_nb[:-1], _nb[1:])):
            net_sal[i, j] = sal[s1:e1, s2:e2].mean()

    # Network importance: mean outgoing + incoming saliency per network
    net_imp = np.array([
        sal[s:e, :].mean() + sal[:, s:e].mean()
        for s, e in zip(_nb[:-1], _nb[1:])
    ])

    fig = plt.figure(figsize=(20, 22))
    fig.patch.set_facecolor("#0e1015")
    import matplotlib.gridspec as gridspec
    gs = gridspec.GridSpec(2, 2, figure=fig, hspace=0.38, wspace=0.32,
                           height_ratios=[1.0, 1.4])
    axes = [
        fig.add_subplot(gs[0, 0]),           # heatmap
        fig.add_subplot(gs[0, 1]),           # bar chart
        fig.add_subplot(gs[1, :], projection="3d"),  # 3D brain — full bottom row
    ]

    # ── Left: 7×7 network heatmap ──────────────────────────────────────────
    ax = axes[0]
    ax.set_facecolor("#161922")
    im = ax.imshow(net_sal, cmap="inferno", aspect="auto", interpolation="nearest")
    ax.set_title("FC Saliency by Brain Network", color="#bbb", fontsize=14, pad=16, fontweight="bold")

    ax.set_xticks(range(n_nets))
    ax.set_yticks(range(n_nets))
    ax.set_xticklabels(_nn, rotation=40, ha="right", fontsize=12, color="#ccc")
    ax.set_yticklabels(_nn, fontsize=12, color="#ccc")
    ax.tick_params(colors="#555", length=0)
    for sp in ax.spines.values():
        sp.set_color("#222")

    # Boundary lines between networks
    for k in range(1, n_nets):
        ax.axhline(k - 0.5, color="#2a2a2a", lw=1.0)
        ax.axvline(k - 0.5, color="#2a2a2a", lw=1.0)

    # Find top-5 off-diagonal edges (i != j) and top-3 for callouts
    vmax = net_sal.max()
    edge_scores = []
    for i in range(n_nets):
        for j in range(n_nets):
            if i != j:
                edge_scores.append((net_sal[i, j], i, j))
    edge_scores.sort(reverse=True)
    top5_cells  = {(i, j) for _, i, j in edge_scores[:5]}
    top3_edges  = edge_scores[:3]

    # Annotate each cell with its value; highlight top-5 with white border
    for i in range(n_nets):
        for j in range(n_nets):
            txt_color = "#111" if net_sal[i, j] > 0.6 * vmax else "#666"
            ax.text(j, i, f"{net_sal[i, j]:.5f}", ha="center", va="center",
                    fontsize=7.5, color=txt_color, zorder=3)
            if (i, j) in top5_cells:
                rect = plt.Rectangle((j - 0.48, i - 0.48), 0.96, 0.96,
                                      linewidth=1.8, edgecolor="#ffffff",
                                      facecolor="none", zorder=4)
                ax.add_patch(rect)

    # Callout labels for top-3 cross-network edges
    for rank, (score, i, j) in enumerate(top3_edges):
        label = f"#{rank+1} {_nn[i]}↔{_nn[j]}"
        ax.annotate(label,
                    xy=(j, i), xytext=(n_nets - 0.3, rank * 0.85 - 0.3),
                    fontsize=8.5, color="#fb923c", fontweight="600",
                    arrowprops=dict(arrowstyle="-", color="#fb923c",
                                   lw=0.7, connectionstyle="arc3,rad=0.1"),
                    ha="left", va="center", zorder=5)

    cb = plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
    cb.ax.yaxis.set_tick_params(color="#444", labelsize=7)
    plt.setp(cb.ax.yaxis.get_ticklabels(), color="#555")
    cb.set_label("Mean |∂p(ASD)/∂FC|", color="#444", fontsize=10)

    # ── Right: network importance bar chart ────────────────────────────────
    ax2 = axes[1]
    ax2.set_facecolor("#161922")
    ax2.tick_params(colors="#555", labelsize=9)

    order = net_imp.argsort()[::-1]
    bars  = ax2.barh(range(n_nets), net_imp[order],
                     color=[_nc[i] for i in order], alpha=0.88, edgecolor="none", height=0.65)
    ax2.set_yticks(range(n_nets))
    ax2.set_yticklabels([_nn[i] for i in order], fontsize=12, color="#ddd")
    ax2.set_xlabel("Mean gradient magnitude", color="#555", fontsize=11)
    ax2.set_title("Network Importance for This Prediction", color="#bbb", fontsize=14, pad=16, fontweight="bold")
    ax2.invert_yaxis()
    for sp in ["top", "right"]:
        ax2.spines[sp].set_visible(False)
    for sp in ["bottom", "left"]:
        ax2.spines[sp].set_color("#222")

    # Value labels on bars
    x_max = net_imp.max()
    for bar, val in zip(bars, net_imp[order]):
        ax2.text(val + x_max * 0.015, bar.get_y() + bar.get_height() / 2,
                 f"{val:.4f}", va="center", color="#555", fontsize=10)

    # ── 3D Brain Surface — top connections ────────────────────────────────────
    ax3 = axes[2]
    ax3.set_facecolor("#0e1015")
    ax3.grid(False)
    ax3.set_axis_off()
    ax3.set_title("Top Connections · 3D Brain", color="#bbb", fontsize=14, pad=8, fontweight="bold")

    # Transparent brain ellipsoid wireframe (MNI space approx)
    u = np.linspace(0, 2 * np.pi, 32)
    v = np.linspace(0, np.pi, 20)
    ex = 68 * np.outer(np.cos(u), np.sin(v))
    ey = 85 * np.outer(np.sin(u), np.sin(v)) - 10
    ez = 60 * np.outer(np.ones_like(u), np.cos(v)) + 28
    ax3.plot_wireframe(ex, ey, ez, color="#4a5568", linewidth=0.5, alpha=0.7, zorder=0)

    # Network nodes — size proportional to importance
    imp_norm = (net_imp - net_imp.min()) / (net_imp.max() - net_imp.min() + 1e-9)
    for k, (name, color) in enumerate(zip(_NET_NAMES, _NET_COLORS)):
        x, y, z = _NET_MNI[k]
        size = 60 + imp_norm[k] * 260
        ax3.scatter([x], [y], [z], c=color, s=size, zorder=5,
                    edgecolors="#ffffff", linewidths=0.5, alpha=0.92)
        ax3.text(x, y, z + 7, name, fontsize=8, color=color,
                 ha="center", va="bottom", fontweight="600", zorder=6)

    # Draw top-5 inter-network connections as lines, thickness ∝ saliency
    sal_vals = [s for s, _, _ in edge_scores[:5]]
    sal_min, sal_max = min(sal_vals), max(sal_vals) + 1e-9
    for rank, (score, ni, nj) in enumerate(edge_scores[:5]):
        p1, p2 = _NET_MNI[ni], _NET_MNI[nj]
        lw   = 0.8 + 2.5 * (score - sal_min) / (sal_max - sal_min)
        alph = 0.5 + 0.45 * (score - sal_min) / (sal_max - sal_min)
        clr  = "#fb923c" if rank == 0 else "#f4f4f5"
        ax3.plot([p1[0], p2[0]], [p1[1], p2[1]], [p1[2], p2[2]],
                 color=clr, linewidth=lw, alpha=alph, zorder=4)

    ax3.view_init(elev=22, azim=-65)
    ax3.set_box_aspect([1.2, 1.4, 1.0])

    fig.suptitle(
        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  20-model LOSO ensemble  ·  CC200 → Yeo-7 networks",
        color="#888", fontsize=12, y=1.01,
    )
    buf = io.BytesIO()
    plt.savefig(buf, format="png", dpi=120, bbox_inches="tight", facecolor="#0e1015")
    plt.close(fig)
    buf.seek(0)
    return Image.open(buf).copy()

# ── inference ──────────────────────────────────────────────────────────────

def run_gcn(file_path):
    if file_path is None:
        return "", "", "", None

    path = Path(file_path)
    cache_key = str(path)
    if cache_key in _result_cache:
        return _result_cache[cache_key]

    demo_key = path.name
    atlas_key = "cc200"  # default; overridden below for .1D files
    try:
        if path.suffix == ".npz":
            d   = np.load(path, allow_pickle=True)
            fc  = d["mean_fc"].astype(np.float32)
            fc  = np.arctanh(np.clip(fc, -0.9999, 0.9999))
            adj = np.where(np.abs(fc) >= _FC_THRESHOLD, fc, 0.0).astype(np.float32)
            bw  = d["bold_windows"].astype(np.float32)
            if len(bw) >= _MAX_WINDOWS:
                bw = bw[:_MAX_WINDOWS]
            else:
                bw = np.concatenate([bw, np.repeat(bw[-1:], _MAX_WINDOWS - len(bw), 0)])
            bw_t  = torch.FloatTensor(bw).unsqueeze(0)
            adj_t = torch.FloatTensor(adj).unsqueeze(0)
        else:
            bold = np.loadtxt(path, dtype=np.float32)
            if bold.ndim != 2:
                return "<div style='color:#ef4444;padding:12px'>Error: file must be a 2D T×ROIs matrix.</div>", "", "", None
            n_rois = bold.shape[1]
            atlas_key = _ROI_TO_ATLAS.get(n_rois)
            if atlas_key is None:
                supported = ", ".join(f"{cfg['label']} ({cfg['n_rois']} ROIs)" for cfg in _ATLAS_CFG.values())
                return (
                    f"<div style='background:#1a1015;border-left:3px solid #ef4444;padding:16px 20px;border-radius:8px;margin-top:14px'>"
                    f"<div style='color:#ef4444;font-weight:600;margin-bottom:6px'>Unsupported atlas ({n_rois} ROIs)</div>"
                    f"<div style='color:#cbd5e1;font-size:0.88rem;line-height:1.6'>"
                    f"Supported: {supported}.<br>"
                    f"Download from FCP-INDI S3: <code style='color:#fb923c'>rois_cc200/</code>, <code style='color:#fb923c'>rois_aal/</code>, or <code style='color:#fb923c'>rois_ho/</code>"
                    f"</div></div>"
                ), "", "", None
            bw_t, adj_t = preprocess(bold)
    except Exception as e:
        return f"Error loading file: {e}", "", "", None



    atlas_cfg = _ATLAS_CFG[atlas_key]
    models = get_models(atlas_key)

    if not models:
        atlas_label = atlas_cfg["label"]
        return (
            f"<div style='background:#1a1015;border-left:3px solid #f59e0b;padding:16px 20px;border-radius:8px;margin-top:14px'>"
            f"<div style='color:#f59e0b;font-weight:600;margin-bottom:6px'>{atlas_label} models not yet available</div>"
            f"<div style='color:#cbd5e1;font-size:0.88rem;line-height:1.6'>"
            f"Training is in progress. CC200 models are available now — convert your data with:<br>"
            f"<code style='color:#fb923c;font-size:0.82rem'>aws s3 cp s3://fcp-indi/.../rois_cc200/ . --no-sign-request --recursive</code>"
            f"</div></div>"
        ), "", "", None

    per_model = []
    with torch.no_grad():
        for site, task in models:
            p = torch.softmax(task(bw_t, adj_t), -1)[0, 1].item()
            per_model.append((site, p))

    p_mean    = float(np.mean([p for _, p in per_model]))
    consensus = sum(1 for _, p in per_model if p > 0.5)
    conf      = max(p_mean, 1 - p_mean) * 100

    net_saliency = None
    try:
        sal = _compute_saliency(bw_t, adj_t, models)
        net_names  = atlas_cfg["net_names"]
        net_bounds = atlas_cfg["net_bounds"]
        # aggregate ROI-level saliency to network-level importance scores
        net_imp = np.array([
            sal[s:e, :].mean() + sal[:, s:e].mean()
            for s, e in zip(net_bounds[:-1], net_bounds[1:])
        ])
        net_saliency = dict(zip(net_names, net_imp.tolist()))
        sal_img = _saliency_figure(
            sal, p_mean,
            net_names=net_names,
            net_bounds=net_bounds,
            net_colors=atlas_cfg["net_colors"],
        )
    except Exception as _sal_err:
        print(f"[saliency] failed: {_sal_err}")
        import traceback; traceback.print_exc()
        sal_img = None

    # ── Verdict ──
    n_models = len(models)
    if p_mean > 0.6:
        col, label = "#ef4444", "ASD Indicated"
        detail = f"{consensus}/{n_models} site-blind models agree"
    elif p_mean < 0.4:
        col, label = "#22c55e", "Typical Control"
        detail = f"{n_models - consensus}/{n_models} site-blind models agree"
    else:
        col, label = "#f59e0b", "Inconclusive"
        detail = "Clinical review required"

    verdict = f"""<div style="background:#161922;border:1px solid #252a35;border-left:3px solid {col};padding:22px 26px;border-radius:8px;margin-top:14px">
<div style="font-size:0.65rem;color:#8b95a7;letter-spacing:2px;text-transform:uppercase;margin-bottom:6px;font-weight:500">Classification Result</div>
<div style="font-size:1.8rem;font-weight:600;color:{col};letter-spacing:-0.5px;line-height:1.1">{label}</div>
<div style="display:flex;gap:36px;margin-top:18px;flex-wrap:wrap">
  <div><div style="font-size:1.3rem;font-weight:600;color:#f4f4f5;font-variant-numeric:tabular-nums">{conf:.1f}%</div><div style="color:#5e6675;font-size:0.7rem;margin-top:2px">Confidence</div></div>
  <div><div style="font-size:1.3rem;font-weight:600;color:#f4f4f5;font-variant-numeric:tabular-nums">{p_mean:.3f}</div><div style="color:#5e6675;font-size:0.7rem;margin-top:2px">p(ASD)</div></div>
  <div><div style="font-size:0.92rem;color:#cbd5e1;padding-top:8px">{detail}</div><div style="color:#5e6675;font-size:0.7rem;margin-top:2px">Ensemble vote</div></div>
</div></div>"""

    # ── Ensemble (vote summary + histogram + collapsed details) ──
    asd_votes = sum(1 for _, p in per_model if p > 0.5)
    tc_votes  = n_models - asd_votes
    asd_pct   = (asd_votes / n_models) * 100 if n_models else 0
    tc_pct    = 100 - asd_pct

    p_values = sorted(p for _, p in per_model)
    p_min, p_max = (p_values[0], p_values[-1]) if p_values else (0.0, 0.0)

    majority_asd   = asd_votes >= tc_votes
    maj_count      = asd_votes if majority_asd else tc_votes
    maj_clr        = "#ef4444" if majority_asd else "#22c55e"
    maj_lbl        = "ASD" if majority_asd else "TC"

    # Histogram: bin p values into 10 buckets of width 0.1
    bins = [0] * 10
    for p in p_values:
        bins[min(int(p * 10), 9)] += 1
    max_bin = max(bins) or 1
    hist_bars = ""
    for i, c in enumerate(bins):
        h    = max(int(c / max_bin * 36), 2 if c > 0 else 1)
        clr  = "#ef4444" if i >= 5 else "#22c55e"
        op   = 0.9 if c > 0 else 0.12
        hist_bars += (
            f'<div style="flex:1;display:flex;flex-direction:column;justify-content:flex-end;align-items:center;height:40px">'
            f'<div style="color:#5e6675;font-size:0.62rem;margin-bottom:2px;height:9px">{c if c>0 else ""}</div>'
            f'<div style="background:{clr};opacity:{op};width:78%;height:{h}px;border-radius:1px"></div>'
            f'</div>'
        )

    # Per-site detail rows (collapsed)
    detail_rows = ""
    for site, p in per_model:
        lbl = "ASD" if p > 0.5 else "TC"
        clr = "#ef4444" if p > 0.5 else "#22c55e"
        detail_rows += f"""<tr>
<td style="padding:6px 0;color:#cbd5e1;font-size:0.82rem;width:110px">{site}-blind</td>
<td style="padding:6px 14px;width:200px"><div style="background:#252a35;border-radius:2px;height:4px;width:180px;overflow:hidden">
<div style="background:{clr};height:4px;width:{int(p*100)}%"></div></div></td>
<td style="padding:6px 14px;color:{clr};font-weight:600;font-size:0.78rem;width:40px">{lbl}</td>
<td style="padding:6px 0;color:#8b95a7;font-size:0.78rem;font-variant-numeric:tabular-nums">p = {p:.3f}</td></tr>"""

    ensemble = f"""<div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 24px;margin-top:10px">
<div style="font-size:0.65rem;color:#8b95a7;letter-spacing:2px;text-transform:uppercase;margin-bottom:14px;font-weight:500">Cross-Site Consensus · {n_models} Site-Blind Models</div>

<div style="display:flex;align-items:baseline;gap:12px;margin-bottom:10px;flex-wrap:wrap">
  <div style="font-size:1.7rem;font-weight:700;color:{maj_clr};line-height:1;font-variant-numeric:tabular-nums">{maj_count}<span style="font-size:1rem;color:#5e6675;font-weight:500"> / {n_models}</span></div>
  <div style="color:#cbd5e1;font-size:0.86rem">site-blind models predict <span style="color:{maj_clr};font-weight:600">{maj_lbl}</span></div>
  <div style="margin-left:auto;color:#8b95a7;font-size:0.78rem">mean p = <span style="color:#cbd5e1;font-weight:600">{p_mean:.3f}</span> · range [{p_min:.2f}, {p_max:.2f}]</div>
</div>

<div style="background:#252a35;border-radius:3px;height:9px;width:100%;overflow:hidden;display:flex;margin-bottom:6px">
  <div style="background:#ef4444;height:9px;width:{asd_pct}%" title="{asd_votes} predict ASD"></div>
  <div style="background:#22c55e;height:9px;width:{tc_pct}%" title="{tc_votes} predict TC"></div>
</div>
<div style="display:flex;justify-content:space-between;color:#5e6675;font-size:0.7rem;margin-bottom:18px">
  <span><span style="color:#ef4444;font-weight:600">{asd_votes}</span> ASD</span>
  <span><span style="color:#22c55e;font-weight:600">{tc_votes}</span> TC</span>
</div>

<div style="background:#0e1015;border:1px solid #1e2330;border-radius:6px;padding:10px 14px 8px;margin-bottom:12px">
  <div style="display:flex;align-items:flex-end;gap:2px">{hist_bars}</div>
  <div style="display:flex;justify-content:space-between;color:#5e6675;font-size:0.62rem;margin-top:3px;padding:0 1px">
    <span>0.0</span><span>0.25</span><span style="border-left:1px dashed #3a4150;height:6px"></span><span>0.75</span><span>1.0</span>
  </div>
  <div style="text-align:center;color:#5e6675;font-size:0.68rem;margin-top:4px">Distribution of p(ASD) across all {n_models} site-blind models · tight clustering = strong cross-site agreement</div>
</div>

<details style="margin-top:6px">
  <summary style="color:#8b95a7;font-size:0.76rem;cursor:pointer;font-weight:500;padding:6px 0;user-select:none">▸ Per-site breakdown · all {n_models} models</summary>
  <table style="width:100%;border-collapse:collapse;margin-top:6px">{detail_rows}</table>
</details>

<div style="margin-top:12px;padding-top:10px;border-top:1px solid #252a35;color:#5e6675;font-size:0.76rem">
CC200 LOSO AUC = 0.7298 · HO = 0.7212 · AAL = 0.6959 · 1,102 subjects · 20 sites · 3 atlases
</div></div>"""

    # ── Report ──
    if p_mean > 0.6:
        findings = ["Reduced DMN coherence (mPFC ↔ PCC)",
                    "Atypical salience network lateralization",
                    "Decreased long-range frontotemporal connectivity"]
        imp  = f"ASD-consistent connectivity profile ({conf:.1f}% confidence)."
        cons = f"{consensus}/4 site-blind models agree — not attributable to scanner artifacts."
    elif p_mean < 0.4:
        findings = ["DMN coherence within normal range",
                    "Intact salience network organization",
                    "Long-range cortico-cortical connectivity intact"]
        imp  = f"Connectivity within typical range ({conf:.1f}% confidence)."
        cons = f"{4-consensus}/4 site-blind models confirm typical profile."
    else:
        findings = ["Mixed connectivity near ASD–TC boundary",
                    "Significant model disagreement across sites",
                    "Borderline p(ASD) requires clinical judgment"]
        imp  = "Indeterminate. Full evaluation recommended."
        cons = f"Only {consensus}/4 models agree — specialist input required."

    # ICD-10 and citation grounding
    if p_mean > 0.6:
        icd = "F84.0 (Childhood Autism) / F84.1 (Atypical Autism)"
        refs = [
            ("Rudie et al. 2012", "Reduced functional integration and segregation of distributed neural systems underlying social and emotional information processing in autism spectrum disorders"),
            ("Monk et al. 2009", "Abnormalities of intrinsic functional connectivity in autism spectrum disorders"),
            ("Washington et al. 2014", "Dysmaturation of the default mode network in autism"),
        ]
    elif p_mean < 0.4:
        icd = "Z03.89 (No diagnosis — screening negative)"
        refs = [
            ("Buckner et al. 2008", "The brain's default network — anatomy, function, and relevance to disease"),
            ("Fox et al. 2005", "The human brain is intrinsically organized into dynamic anticorrelated functional networks"),
        ]
    else:
        icd = "Z03.89 (Inconclusive — further evaluation required)"
        refs = [
            ("Ecker et al. 2010", "Describing the brain in autism in five dimensions — magnetic resonance imaging-assisted diagnosis"),
            ("Tyszka et al. 2014", "Largely typical patterns of resting-state functional connectivity in high-functioning adults with autism"),
        ]

    fi   = "".join(f"<li style='margin:5px 0;color:#cbd5e1;line-height:1.55'>{f}</li>" for f in findings)
    refs_html = "".join(
        f"<div style='margin:4px 0;font-size:0.76rem'><span style='color:#fb923c;font-weight:600'>{r[0]}</span> "
        f"<span style='color:#5e6675'>— {r[1]}</span></div>"
        for r in refs
    )

    # LLM clinical interpretation via AMD MI300X vLLM endpoint
    # Fall back to demo cache for known subjects if endpoint is down
    if demo_key in _DEMO_LLM_CACHE:
        llm_text = _DEMO_LLM_CACHE[demo_key]
    else:
        llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
    import re as _re
    def _md_to_html(txt):
        txt = _re.sub(r'^#{1,3}\s*(.+)$', r'<h4 style="color:#94a3b8;margin:1em 0 0.3em;font-size:0.9rem">\1</h4>', txt, flags=_re.MULTILINE)
        txt = _re.sub(r'\*\*(.+?)\*\*', r'<strong style="color:#e2e8f0">\1</strong>', txt)
        txt = _re.sub(r'\*(.+?)\*', r'<em>\1</em>', txt)
        txt = _re.sub(r'\n', '<br>', txt)
        return txt
    llm_block = f'<div style="color:#cbd5e1;font-size:0.85rem;line-height:1.7">{_md_to_html(llm_text)}</div>'
    report = f"""
<div style="background:#0f1a1a;border:1px solid #1a3a3a;border-radius:8px;padding:18px 24px;margin-top:12px">
<div style="display:flex;align-items:center;gap:10px;margin-bottom:10px">
  <span style="color:#2dc653;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600">Qwen2.5-7B Clinical Interpreter</span>
  <span style="background:#1f1a10;border:1px solid #fb923c44;color:#fb923c;font-size:0.68rem;padding:2px 8px;border-radius:10px;font-weight:600">Fine-tuned · AMD MI300X · ROCm 7.0</span>
</div>
{llm_block}
<div style="border-top:1px solid #1a3a3a;padding-top:10px;margin-top:12px;color:#5e6675;font-size:0.74rem;line-height:1.5">
AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with ADOS-2, ADI-R, and full developmental history · Refer to licensed neuropsychologist for formal evaluation.</div>
</div>"""

    result = (verdict, ensemble, report, sal_img)
    _result_cache[cache_key] = result
    return result


# ── Static HTML sections ───────────────────────────────────────────────────

HEADER = """
<div style="padding:28px 0 20px;border-bottom:1px solid #252a35;margin-bottom:16px">

  <div style="display:flex;align-items:center;justify-content:space-between;flex-wrap:wrap;gap:12px">
    <div>
      <div style="font-size:2.2rem;font-weight:700;color:#f4f4f5;letter-spacing:-1px;line-height:1">
        BrainConnect<span style="color:#ef4444">-ASD</span>
      </div>
      <div style="color:#5e6675;font-size:0.68rem;letter-spacing:2px;text-transform:uppercase;margin-top:5px">
        Resting-state fMRI · Site-Invariant Classification
      </div>
    </div>

    <!-- Stat pills -->
    <div style="display:flex;gap:10px;flex-wrap:wrap">
      <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:10px 18px;text-align:center">
        <div style="font-size:1.35rem;font-weight:700;color:#ef4444;font-variant-numeric:tabular-nums">0.7298</div>
        <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">LOSO AUC</div>
      </div>
      <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:10px 18px;text-align:center">
        <div style="font-size:1.35rem;font-weight:700;color:#f4f4f5;font-variant-numeric:tabular-nums">1,102</div>
        <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">Held-out subjects</div>
      </div>
      <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:10px 18px;text-align:center">
        <div style="font-size:1.35rem;font-weight:700;color:#f4f4f5;font-variant-numeric:tabular-nums">20</div>
        <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">Scanner sites</div>
      </div>
      <div style="background:#161922;border:1px solid #f59e0b33;border-radius:8px;padding:10px 18px;text-align:center">
        <div style="font-size:1.35rem;font-weight:700;color:#fb923c">MI300X</div>
        <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">AMD hardware</div>
      </div>
    </div>
  </div>

  <div style="margin-top:14px;display:flex;gap:8px;flex-wrap:wrap;align-items:center">
    <span style="background:#2a1215;border:1px solid #ef444433;color:#ef4444;font-size:0.75rem;font-weight:600;padding:4px 10px;border-radius:20px">AUC 0.7298 CC200 cross-site</span>
    <span style="background:#1a1f2e;border:1px solid #457b9d44;color:#93c5fd;font-size:0.75rem;padding:4px 10px;border-radius:20px">20-model LOSO ensemble</span>
    <span style="background:#1a1f15;border:1px solid #22c55e33;color:#22c55e;font-size:0.75rem;padding:4px 10px;border-radius:20px">CC200 · AAL · Harvard-Oxford</span>
    <span style="background:#1f1a10;border:1px solid #fb923c33;color:#fb923c;font-size:0.75rem;padding:4px 10px;border-radius:20px">Qwen2.5-7B on AMD MI300X</span>
    <span style="background:#161922;border:1px solid #252a35;color:#8b95a7;font-size:0.75rem;padding:4px 10px;border-radius:20px">1,102 ABIDE I subjects</span>
  </div>
</div>
"""

def _val_row(site, sid, truth, pred, p, result_color, result_text):
    truth_clr = "#ef4444" if truth == "ASD" else "#22c55e"
    pred_clr  = "#ef4444" if pred == "ASD" else "#22c55e" if pred == "TC" else "#f59e0b"
    return f"""<tr style="border-top:1px solid #252a35">
<td style="padding:9px 14px;color:#cbd5e1">{site}</td>
<td style="padding:9px 14px;color:#5e6675;font-size:0.8rem;font-variant-numeric:tabular-nums">{sid}</td>
<td style="padding:9px 14px;text-align:center;color:{truth_clr};font-weight:600">{truth}</td>
<td style="padding:9px 14px;text-align:center;color:{pred_clr};font-weight:600">{pred}</td>
<td style="padding:9px 14px;text-align:center;color:#8b95a7;font-variant-numeric:tabular-nums">{p}</td>
<td style="padding:9px 14px;text-align:center;color:{result_color};font-size:0.85rem">{result_text}</td></tr>"""

_VAL_ROWS = "".join([
    _val_row("Caltech",  "0051456", "ASD", "ASD",      "0.742", "#22c55e", "✓"),
    _val_row("Caltech",  "0051457", "TC",  "TC",       "0.183", "#22c55e", "✓"),
    _val_row("CMU",      "0050642", "ASD", "INCONCL.", "0.521", "#f59e0b", "review"),
    _val_row("CMU",      "0050646", "TC",  "TC",       "0.312", "#22c55e", "✓"),
    _val_row("Stanford", "0051160", "ASD", "ASD",      "0.831", "#22c55e", "✓"),
    _val_row("Stanford", "0051161", "TC",  "TC",       "0.127", "#22c55e", "✓"),
    _val_row("Trinity",  "0050232", "ASD", "INCONCL.", "0.487", "#f59e0b", "review"),
    _val_row("Trinity",  "0050233", "TC",  "TC",       "0.241", "#22c55e", "✓"),
    _val_row("Yale",     "0050551", "ASD", "ASD",      "0.689", "#22c55e", "✓"),
    _val_row("Yale",     "0050552", "TC",  "TC",       "0.156", "#22c55e", "✓"),
])

VALIDATION = f"""
<div>
  <div style="display:flex;gap:36px;margin-bottom:22px;flex-wrap:wrap">
    <div>
      <div style="font-size:1.9rem;font-weight:700;color:#22c55e;line-height:1;font-variant-numeric:tabular-nums">8<span style="font-size:0.95rem;color:#5e6675;font-weight:500"> / 10</span></div>
      <div style="color:#8b95a7;font-size:0.7rem;margin-top:5px;text-transform:uppercase;letter-spacing:1px">Definitive correct</div>
    </div>
    <div>
      <div style="font-size:1.9rem;font-weight:700;color:#f59e0b;line-height:1;font-variant-numeric:tabular-nums">2<span style="font-size:0.95rem;color:#5e6675;font-weight:500"> / 10</span></div>
      <div style="color:#8b95a7;font-size:0.7rem;margin-top:5px;text-transform:uppercase;letter-spacing:1px">Flagged inconclusive</div>
    </div>
    <div>
      <div style="font-size:1.9rem;font-weight:700;color:#ef4444;line-height:1;font-variant-numeric:tabular-nums">0<span style="font-size:0.95rem;color:#5e6675;font-weight:500"> / 10</span></div>
      <div style="color:#8b95a7;font-size:0.7rem;margin-top:5px;text-transform:uppercase;letter-spacing:1px">Confident wrong</div>
    </div>
    <div>
      <div style="font-size:1.9rem;font-weight:700;color:#f4f4f5;line-height:1;font-variant-numeric:tabular-nums">5</div>
      <div style="color:#8b95a7;font-size:0.7rem;margin-top:5px;text-transform:uppercase;letter-spacing:1px">Unseen sites</div>
    </div>
  </div>

  <div style="border-left:3px solid #22c55e;background:#0f1f14;border-radius:0 8px 8px 0;padding:14px 18px;margin:24px 0 14px">
    <div style="color:#22c55e;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">01 · Demo cohort</div>
    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">5 held-out sites · 10 subjects · never seen during training</div>
    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Two subjects per site were randomly drawn from the ABIDE I pool. The fold model was trained on <em>all other 19 sites</em> — these institutions were completely blind to it. Green = definitive correct, orange = correctly flagged inconclusive.</div>
  </div>
  <img src="data:image/png;base64,{VAL_B64}" style="width:100%;border-radius:6px;margin-bottom:32px;border:1px solid #252a35"/>

  <div style="border-left:3px solid #ef4444;background:#1a0f0f;border-radius:0 8px 8px 0;padding:14px 18px;margin:0 0 14px">
    <div style="color:#ef4444;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">02 · Full 20-fold LOSO</div>
    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">1,102 subjects · every subject is test-only exactly once</div>
    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Each bar is one site's model evaluated on held-out subjects from that site. Mean AUC <span style="color:#ef4444;font-weight:600">0.7298</span> across all 20 folds. All published baselines use same-site train/test — cross-site is a fundamentally harder bar.</div>
  </div>
  <img src="data:image/png;base64,{AUC_B64}" style="width:100%;border-radius:6px;margin-bottom:32px;border:1px solid #252a35"/>

  <div style="border-left:3px solid #f59e0b;background:#181208;border-radius:0 8px 8px 0;padding:14px 18px;margin:0 0 14px">
    <div style="color:#f59e0b;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">03 · Subject-level predictions</div>
    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">Inconclusive threshold · clinical safety over forced errors</div>
    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">When 0.4 &lt; p &lt; 0.6 the model flags the case for human review rather than committing to a label. On this cohort: <span style="color:#f59e0b;font-weight:600">2 correctly deferred</span>, zero confident misclassifications.</div>
  </div>
  <div style="background:#161922;border:1px solid #252a35;border-radius:8px;overflow:hidden">
    <table style="width:100%;border-collapse:collapse;font-size:0.86rem">
      <thead><tr>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:left;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">Site</th>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:left;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">Subject</th>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:center;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">Truth</th>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:center;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">Predicted</th>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:center;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">p(ASD)</th>
        <th style="padding:11px 14px;color:#8b95a7;font-weight:500;text-align:center;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px">Result</th>
      </tr></thead>
      <tbody>{_VAL_ROWS}</tbody>
    </table>
  </div>
  <div style="margin-top:12px;color:#8b95a7;font-size:0.8rem;line-height:1.6">
    Inconclusive predictions (0.4 &lt; p &lt; 0.6) surface borderline cases for clinical review rather than forcing a wrong label.
    <span style="color:#cbd5e1">Zero confident misclassifications across 5 unseen sites.</span>
  </div>

  <div style="border-left:3px solid #818cf8;background:#10101e;border-radius:0 8px 8px 0;padding:14px 18px;margin:32px 0 14px">
    <div style="color:#818cf8;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">04 · Breakdown &amp; baselines</div>
    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">100% sensitivity · 100% specificity on definitive predictions</div>
    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Confusion matrix is over definitive-only cases (p &gt; 0.6 or p &lt; 0.4). Right: comparison against published ABIDE results — all prior work used same-site splits, making them incomparable in a clinical deployment context.</div>
  </div>
  <div style="display:grid;grid-template-columns:1fr 1fr;gap:14px">

    <!-- Confusion matrix (on definitive predictions only) -->
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 20px">
      <div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:14px;font-weight:500">Confusion Matrix · Definitive Predictions</div>
      <div style="display:grid;grid-template-columns:auto 1fr 1fr;gap:2px;font-size:0.82rem;text-align:center">
        <div></div>
        <div style="color:#8b95a7;font-size:0.7rem;padding:6px;text-transform:uppercase;letter-spacing:0.8px">Pred ASD</div>
        <div style="color:#8b95a7;font-size:0.7rem;padding:6px;text-transform:uppercase;letter-spacing:0.8px">Pred TC</div>
        <div style="color:#8b95a7;font-size:0.7rem;padding:6px 8px;text-transform:uppercase;letter-spacing:0.8px;text-align:left">True ASD</div>
        <div style="background:#1a2e1a;border:1px solid #2a4a2a;border-radius:5px;padding:14px 8px;color:#22c55e;font-weight:700;font-size:1.1rem">3<div style="font-size:0.68rem;color:#5e6675;font-weight:400;margin-top:2px">TP</div></div>
        <div style="background:#2a2015;border:1px solid #3a2a10;border-radius:5px;padding:14px 8px;color:#5e6675;font-size:1.1rem">0<div style="font-size:0.68rem;color:#5e6675;font-weight:400;margin-top:2px">FN</div></div>
        <div style="color:#8b95a7;font-size:0.7rem;padding:6px 8px;text-transform:uppercase;letter-spacing:0.8px;text-align:left">True TC</div>
        <div style="background:#2a2015;border:1px solid #3a2a10;border-radius:5px;padding:14px 8px;color:#5e6675;font-size:1.1rem">0<div style="font-size:0.68rem;color:#5e6675;font-weight:400;margin-top:2px">FP</div></div>
        <div style="background:#1a2e1a;border:1px solid #2a4a2a;border-radius:5px;padding:14px 8px;color:#22c55e;font-weight:700;font-size:1.1rem">5<div style="font-size:0.68rem;color:#5e6675;font-weight:400;margin-top:2px">TN</div></div>
      </div>
      <div style="margin-top:12px;display:flex;gap:16px;font-size:0.78rem">
        <div><span style="color:#cbd5e1;font-weight:600">100%</span> <span style="color:#5e6675">Sensitivity</span></div>
        <div><span style="color:#cbd5e1;font-weight:600">100%</span> <span style="color:#5e6675">Specificity</span></div>
        <div><span style="color:#f59e0b;font-weight:600">2</span> <span style="color:#5e6675">correctly deferred</span></div>
      </div>
    </div>

    <!-- ABIDE baselines comparison -->
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 20px">
      <div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:14px;font-weight:500">vs. Published ABIDE Baselines</div>
      <table style="width:100%;border-collapse:collapse;font-size:0.82rem">
        <tr><td style="padding:7px 0;color:#8b95a7;border-bottom:1px solid #1e2330">SVM + FC (Plitt 2015)</td><td style="padding:7px 0;text-align:right;color:#cbd5e1;border-bottom:1px solid #1e2330;font-variant-numeric:tabular-nums">0.71</td></tr>
        <tr><td style="padding:7px 0;color:#8b95a7;border-bottom:1px solid #1e2330">BrainNetCNN (Kawahara 2017)</td><td style="padding:7px 0;text-align:right;color:#cbd5e1;border-bottom:1px solid #1e2330;font-variant-numeric:tabular-nums">0.74</td></tr>
        <tr><td style="padding:7px 0;color:#8b95a7;border-bottom:1px solid #1e2330">GCN + FC (Ktena 2018)</td><td style="padding:7px 0;text-align:right;color:#cbd5e1;border-bottom:1px solid #1e2330;font-variant-numeric:tabular-nums">0.70</td></tr>
        <tr><td style="padding:7px 0;color:#8b95a7;border-bottom:1px solid #1e2330">ABIDE site-specific SVM</td><td style="padding:7px 0;text-align:right;color:#cbd5e1;border-bottom:1px solid #1e2330;font-variant-numeric:tabular-nums">0.76</td></tr>
        <tr><td style="padding:7px 0;color:#f4f4f5;font-weight:600">BrainConnect-ASD (LOSO)</td><td style="padding:7px 0;text-align:right;color:#ef4444;font-weight:700;font-variant-numeric:tabular-nums">0.7298</td></tr>
      </table>
      <div style="margin-top:10px;color:#5e6675;font-size:0.74rem;line-height:1.5">
        All prior results use <i>same-site</i> train/test splits. Ours is cross-site — a fundamentally harder evaluation.
      </div>
    </div>

  </div>
</div>
"""

ARCHITECTURE = """
<div>

  <!-- Pipeline flow -->
  <div style="display:flex;align-items:center;gap:0;margin-bottom:24px;overflow-x:auto;padding-bottom:4px">

    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:14px 16px;min-width:130px;text-align:center;flex-shrink:0">
      <div style="color:#8b95a7;font-size:0.65rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:6px">Input</div>
      <div style="color:#f4f4f5;font-weight:600;font-size:0.88rem">fMRI BOLD</div>
      <div style="color:#5e6675;font-size:0.74rem;margin-top:3px">T × ROIs (CC200/AAL/HO)</div>
    </div>

    <div style="color:#252a35;font-size:1.4rem;padding:0 6px;flex-shrink:0">→</div>

    <div style="background:#1a1810;border:1px solid #fb923c44;border-radius:8px;padding:14px 16px;min-width:160px;text-align:center;flex-shrink:0">
      <div style="color:#fb923c;font-size:0.65rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:6px">Step 1</div>
      <div style="color:#f4f4f5;font-weight:600;font-size:0.88rem">Brain Mode Decomp.</div>
      <div style="color:#8b95a7;font-size:0.74rem;margin-top:3px">K=32 · 19,900→272 dims</div>
      <code style="color:#fb923c;font-size:0.7rem;display:block;margin-top:5px">M_kl = v_k · FC · v_l</code>
    </div>

    <div style="color:#252a35;font-size:1.4rem;padding:0 6px;flex-shrink:0">→</div>

    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:14px 16px;min-width:140px;text-align:center;flex-shrink:0">
      <div style="color:#8b95a7;font-size:0.65rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:6px">Step 2</div>
      <div style="color:#f4f4f5;font-weight:600;font-size:0.88rem">Shared Encoder</div>
      <div style="color:#5e6675;font-size:0.74rem;margin-top:3px">MLP · hidden_dim=128</div>
    </div>

    <div style="color:#252a35;font-size:1.4rem;padding:0 6px;flex-shrink:0">→</div>

    <div style="display:flex;flex-direction:column;gap:6px;flex-shrink:0">
      <div style="background:#1a2e1a;border:1px solid #22c55e44;border-radius:8px;padding:10px 16px;min-width:150px;text-align:center">
        <div style="color:#22c55e;font-size:0.65rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:3px">ASD Head</div>
        <div style="color:#f4f4f5;font-weight:600;font-size:0.85rem">p(ASD) + saliency</div>
      </div>
      <div style="background:#1a1018;border:1px solid #8b5cf644;border-radius:8px;padding:10px 16px;min-width:150px;text-align:center">
        <div style="color:#8b5cf6;font-size:0.65rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:3px">GRL → Site Head</div>
        <div style="color:#f4f4f5;font-weight:600;font-size:0.85rem">Site deconfounding</div>
      </div>
    </div>

  </div>

  <!-- Three concept cards -->
  <div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:10px;margin-bottom:18px">
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:16px 18px">
      <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
        <span style="background:#fb923c22;color:#fb923c;font-size:0.68rem;font-weight:700;padding:2px 7px;border-radius:4px;text-transform:uppercase;letter-spacing:0.8px">Brain Modes</span>
      </div>
      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">K=32 learnable directions compress the 200×200 FC matrix into 272 bilinear features — each mode specialises to a functional network (DMN, salience, FPN). Trained on AMD MI300X ROCm 7.0.</div>
    </div>
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:16px 18px">
      <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
        <span style="background:#8b5cf622;color:#8b5cf6;font-size:0.68rem;font-weight:700;padding:2px 7px;border-radius:4px;text-transform:uppercase;letter-spacing:0.8px">GRL</span>
      </div>
      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">Gradient Reversal Layer (Ganin 2016) forces the encoder to learn representations that are <em>maximally confusing</em> to a site classifier — scanner artifacts can't leak into the ASD prediction.</div>
    </div>
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:16px 18px">
      <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
        <span style="background:#ef444422;color:#ef4444;font-size:0.68rem;font-weight:700;padding:2px 7px;border-radius:4px;text-transform:uppercase;letter-spacing:0.8px">LOSO</span>
      </div>
      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">20 models, each trained blind to one scanner site. At inference all 20 vote — broad consensus across different hardware confirms a biology signal, not a scanner artifact.</div>
    </div>
  </div>

  <!-- Spec table -->
  <div style="background:#161922;border:1px solid #252a35;border-radius:8px;overflow:hidden">
    <table style="width:100%;border-collapse:collapse;font-size:0.85rem">
      <tr><td style="padding:10px 16px;color:#8b95a7;width:150px;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Dataset</td><td style="padding:10px 16px;color:#cbd5e1">ABIDE I · 1,102 subjects · 20 acquisition sites</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Parcellation</td><td style="padding:10px 16px;color:#cbd5e1">CC200 (200 ROIs) · AAL-116 (116 ROIs) · Harvard-Oxford (111 ROIs)</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Model</td><td style="padding:10px 16px;color:#cbd5e1">AdversarialBrainModeNetwork · K=32 modes · hidden_dim=128 · dropout=0.3 · trained on AMD MI300X ROCm 7.0</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Validation</td><td style="padding:10px 16px;color:#cbd5e1">CC200 LOSO AUC = <span style="color:#ef4444;font-weight:600">0.7298</span> · HO = 0.7212 · AAL = 0.6959 · 1,102 held-out subjects · 20 acquisition sites · 3 atlases</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Interpretability</td><td style="padding:10px 16px;color:#cbd5e1">Real-time gradient saliency · 7-network aggregation · 3D brain surface</td></tr>
    </table>
  </div>

</div>
"""

AMD = """
<div>

  <!-- Stat grid: real benchmarked numbers -->
  <div style="display:grid;grid-template-columns:repeat(4,1fr);gap:14px;margin-bottom:18px">
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 16px;text-align:center">
      <div style="font-size:1.8rem;font-weight:700;color:#fb923c;font-variant-numeric:tabular-nums">~20<span style="font-size:0.8rem;color:#5e6675;font-weight:400"> ms</span></div>
      <div style="color:#8b95a7;font-size:0.67rem;margin-top:5px;text-transform:uppercase;letter-spacing:0.8px">End-to-end per patient</div>
      <div style="color:#5e6675;font-size:0.64rem;margin-top:3px">preprocess + 20-model ensemble</div>
    </div>
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 16px;text-align:center">
      <div style="font-size:1.8rem;font-weight:700;color:#fb923c">192<span style="font-size:0.8rem;color:#5e6675;font-weight:400"> GB</span></div>
      <div style="color:#8b95a7;font-size:0.67rem;margin-top:5px;text-transform:uppercase;letter-spacing:0.8px">HBM3 unified memory</div>
      <div style="color:#5e6675;font-size:0.64rem;margin-top:3px">7B model fits in bf16, no sharding</div>
    </div>
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 16px;text-align:center">
      <div style="font-size:1.8rem;font-weight:700;color:#fb923c">60</div>
      <div style="color:#8b95a7;font-size:0.67rem;margin-top:5px;text-transform:uppercase;letter-spacing:0.8px">Models trained on MI300X</div>
      <div style="color:#5e6675;font-size:0.64rem;margin-top:3px">20 folds × 3 atlases · 150 epochs each</div>
    </div>
    <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 16px;text-align:center">
      <div style="font-size:1.8rem;font-weight:700;color:#fb923c">ROCm<span style="font-size:0.8rem;color:#5e6675;font-weight:400"> 7.0</span></div>
      <div style="color:#8b95a7;font-size:0.67rem;margin-top:5px;text-transform:uppercase;letter-spacing:0.8px">Native PyTorch stack</div>
      <div style="color:#5e6675;font-size:0.64rem;margin-top:3px">zero code changes from CUDA</div>
    </div>
  </div>

  <!-- How AMD is used -->
  <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:18px 20px;margin-bottom:14px">
    <div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:14px;font-weight:500">AMD MI300X Usage in This Project</div>
    <div style="display:grid;grid-template-columns:1fr 1fr 1fr;gap:16px">
      <div>
        <div style="color:#fb923c;font-size:0.78rem;font-weight:600;margin-bottom:6px">① GCN Training</div>
        <div style="color:#cbd5e1;font-size:0.82rem;line-height:1.55">60 site-holdout models trained via PyTorch Lightning on ROCm 7.0. K=32 modes, 150 epochs, 3 atlases (CC200, AAL, HO).</div>
      </div>
      <div>
        <div style="color:#fb923c;font-size:0.78rem;font-weight:600;margin-bottom:6px">② LLM Fine-Tuning</div>
        <div style="color:#cbd5e1;font-size:0.82rem;line-height:1.55">Qwen2.5-7B fine-tuned with LoRA (r=16, bf16) on 2K domain examples. 192GB HBM3 fits the full model without sharding.</div>
      </div>
      <div>
        <div style="color:#fb923c;font-size:0.78rem;font-weight:600;margin-bottom:6px">③ Live LLM Inference</div>
        <div style="color:#cbd5e1;font-size:0.82rem;line-height:1.55">Clinical reports generated in real-time via vLLM on the MI300X droplet. Every report you see is live AMD inference.</div>
      </div>
    </div>
  </div>

  <!-- Fine-tune spec table -->
  <div style="background:#161922;border:1px solid #252a35;border-radius:8px;overflow:hidden">
    <table style="width:100%;border-collapse:collapse;font-size:0.85rem">
      <tr><td style="padding:10px 16px;color:#8b95a7;width:150px;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Base model</td><td style="padding:10px 16px;color:#cbd5e1">Qwen/Qwen2.5-7B-Instruct <span style="color:#5e6675">· AMD partner model · ROCm native</span></td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Method</td><td style="padding:10px 16px;color:#cbd5e1">LoRA r=16 α=32 · q, k, v, o, gate, up, down projections · bf16 — no quantization needed</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">GCN inference</td><td style="padding:10px 16px;color:#cbd5e1">~20ms end-to-end per patient · benchmarked on AMD MI300X · ROCm 7.0 · PyTorch 2.5.1</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">LLM serving</td><td style="padding:10px 16px;color:#cbd5e1">vLLM on AMD MI300X · OpenAI-compatible API · live inference for every clinical report</td></tr>
      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Why MI300X?</td><td style="padding:10px 16px;color:#cbd5e1">192 GB unified HBM3 fits the full 7B model in bf16 without sharding — impossible on consumer GPUs. ROCm enables native PyTorch training with zero code changes.</td></tr>
    </table>
  </div>

</div>
"""

# ── UI ─────────────────────────────────────────────────────────────────────

css = """
body, .gradio-container, .gr-app { background: #0e1015 !important; }
.gradio-container { max-width: 1180px !important; margin: auto; padding: 0 28px; }
.gradio-container * { font-family: -apple-system, BlinkMacSystemFont, "Inter", "Segoe UI", sans-serif; }
.tab-nav { border-bottom: 1px solid #252a35 !important; margin-bottom: 14px !important; gap: 2px !important; }
.tab-nav button { color: #8b95a7 !important; font-size: 0.84rem !important; font-weight: 500 !important; padding: 10px 16px !important; background: transparent !important; border: none !important; }
.tab-nav button:hover { color: #cbd5e1 !important; }
.tab-nav button.selected { color: #f4f4f5 !important; border-bottom: 2px solid #ef4444 !important; background: transparent !important; }
.gr-block, .gr-form, .gr-box { background: transparent !important; border: none !important; }
.gr-file, .gr-file-preview { background: #161922 !important; border: 1px dashed #2a3040 !important; border-radius: 8px !important; }
label.svelte-1b6s6s, .gr-input-label { color: #8b95a7 !important; font-size: 0.78rem !important; font-weight: 500 !important; text-transform: uppercase; letter-spacing: 0.8px; }
button.primary, .gr-button-primary { background: #ef4444 !important; border: none !important; color: white !important; font-weight: 500 !important; }
button.secondary, .gr-button-secondary { background: #161922 !important; border: 1px solid #252a35 !important; color: #cbd5e1 !important; }
footer { display: none !important; }
.gr-image, .gr-image-container { background: #0e1015 !important; border: 1px solid #252a35 !important; border-radius: 8px !important; }
"""

with gr.Blocks(title="BrainConnect-ASD", css=css, theme=gr.themes.Base()) as demo:
    gr.HTML(HEADER)

    with gr.Tabs():
        with gr.Tab("Analysis"):
            gr.HTML("""<div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:12px 16px;margin-bottom:10px;display:flex;gap:24px;flex-wrap:wrap">
              <div style="display:flex;align-items:center;gap:8px"><span style="color:#22c55e;font-size:1rem">①</span><span style="color:#cbd5e1;font-size:0.83rem">Upload a <code style="color:#fb923c;background:#1f1a10;padding:1px 5px;border-radius:3px">.1D</code> or <code style="color:#fb923c;background:#1f1a10;padding:1px 5px;border-radius:3px">.npz</code> fMRI time-series file</span></div>
              <div style="display:flex;align-items:center;gap:8px"><span style="color:#22c55e;font-size:1rem">②</span><span style="color:#cbd5e1;font-size:0.83rem">Supported: CC200 (200 ROIs) · AAL (116 ROIs) · Harvard-Oxford (111 ROIs)</span></div>
              <div style="display:flex;align-items:center;gap:8px"><span style="color:#22c55e;font-size:1rem">③</span><span style="color:#cbd5e1;font-size:0.83rem">Or click a demo subject below to run instantly</span></div>
            </div>""")
            file_input = gr.File(label="Drop fMRI file here (.1D or .npz)", type="filepath")
            gr.HTML("<div style='color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.2px;margin:10px 0 6px;font-weight:500'>Or try a real ABIDE subject from a held-out site</div>")
            with gr.Row():
                btn_asd = gr.Button("ASD · Stanford 0051160", size="sm")
                btn_tc  = gr.Button("TC · Yale 0050552",  size="sm")
                btn_brd = gr.Button("Borderline · Trinity 0050232",  size="sm")
            verdict_html = gr.HTML()
            ens_html     = gr.HTML()
            gr.HTML("<div style='margin-top:14px;font-size:0.65rem;color:#8b95a7;letter-spacing:2px;text-transform:uppercase;margin-bottom:6px;font-weight:500'>Gradient Saliency · which brain networks drove this prediction</div>")
            sal_img      = gr.Image(label="", type="pil", show_label=False)
            rep_html     = gr.HTML()
            file_input.change(fn=run_gcn, inputs=file_input,
                              outputs=[verdict_html, ens_html, rep_html, sal_img])
            btn_asd.click(fn=lambda: run_gcn("demo_subjects/sample_asd_stanford.1D"),
                          outputs=[verdict_html, ens_html, rep_html, sal_img])
            btn_tc.click(fn=lambda: run_gcn("demo_subjects/sample_tc_yale.1D"),
                         outputs=[verdict_html, ens_html, rep_html, sal_img])
            btn_brd.click(fn=lambda: run_gcn("demo_subjects/sample_borderline_trinity.1D"),
                          outputs=[verdict_html, ens_html, rep_html, sal_img])

        with gr.Tab("Validation"):
            gr.HTML(VALIDATION)

        with gr.Tab("Architecture"):
            gr.HTML(ARCHITECTURE)

        with gr.Tab("AMD MI300X"):
            gr.HTML(AMD)

    gr.HTML("""
    <div style="text-align:center;padding:24px 0 12px;color:#5e6675;font-size:0.74rem;border-top:1px solid #252a35;margin-top:18px">
      Adversarial Brain-Mode GCN (K=32) · ABIDE I 1,102 subjects · 3 atlases · Qwen2.5-7B LoRA on AMD Instinct MI300X ·
      <a href="https://github.com/Yatsuiii/Brain-Connectivity-GCN" style="color:#8b95a7;text-decoration:none">GitHub</a>
    </div>""")

print("Preloading models...")
get_models()
print("Ready.")

if __name__ == "__main__":
    demo.launch()