Upload folder using huggingface_hub

app.py

@@ -17,18 +17,62 @@ _STEP         = 3
 _MAX_WINDOWS  = 30
 _FC_THRESHOLD = 0.2
 
-# CC200 atlas (Craddock 2012) → approximate Yeo 7-network parcellation
-_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 = {
-    "NYU":  Path("checkpoints/nyu.ckpt"),
-    "USM":  Path("checkpoints/usm.ckpt"),
-    "UCLA": Path("checkpoints/ucla.ckpt"),
-    "UM":   Path("checkpoints/um.ckpt"),
+# ── 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": {
+            "NYU":  Path("checkpoints/nyu.ckpt"),
+            "USM":  Path("checkpoints/usm.ckpt"),
+            "UCLA": Path("checkpoints/ucla.ckpt"),
+            "UM":   Path("checkpoints/um.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": {
+            "NYU":  Path("checkpoints/aal_nyu.ckpt"),
+            "USM":  Path("checkpoints/aal_usm.ckpt"),
+            "UCLA": Path("checkpoints/aal_ucla.ckpt"),
+            "UM":   Path("checkpoints/aal_um.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": {
+            "NYU":  Path("checkpoints/ho_nyu.ckpt"),
+            "USM":  Path("checkpoints/ho_usm.ckpt"),
+            "UCLA": Path("checkpoints/ho_ucla.ckpt"),
+            "UM":   Path("checkpoints/ho_um.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):
@@ -60,21 +104,23 @@ def preprocess(bold):
 
 # ── model loading ──────────────────────────────────────────────────────────
 
-_models = None
+_model_cache: dict[str, list] = {}
 
-def get_models():
-    global _models
-    if _models is not None:
-        return _models
+def get_models(atlas: str = "cc200"):
+    global _model_cache
+    if atlas in _model_cache:
+        return _model_cache[atlas]
     from brain_gcn.tasks import ClassificationTask
-    _models = []
-    for site, ckpt in _CKPTS.items():
+    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))
-    return _models
+        models.append((site, task))
+    _model_cache[atlas] = models
+    return models
 
 # ── gradient saliency ──────────────────────────────────────────────────────
 
@@ -99,7 +145,7 @@ _NET_MNI = np.array([
     [ 14,   4,   4],   # Subcortical (thalamus)
 ], dtype=np.float32)
 
-def _saliency_figure(sal, p_mean):
+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
@@ -107,18 +153,21 @@ def _saliency_figure(sal, p_mean):
     from mpl_toolkits.mplot3d.art3d import Line3DCollection
     from PIL import Image
 
-    n_nets = len(_NET_NAMES)
+    _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 200×200 saliency → 7×7 network-level matrix
+    # Aggregate N×N saliency → 7×7 network-level matrix
     net_sal = np.zeros((n_nets, n_nets))
-    for i, (s1, e1) in enumerate(zip(_NET_BOUNDS[:-1], _NET_BOUNDS[1:])):
-        for j, (s2, e2) in enumerate(zip(_NET_BOUNDS[:-1], _NET_BOUNDS[1:])):
+    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(_NET_BOUNDS[:-1], _NET_BOUNDS[1:])
+        for s, e in zip(_nb[:-1], _nb[1:])
     ])
 
     fig = plt.figure(figsize=(18, 5.5))
@@ -137,8 +186,8 @@ def _saliency_figure(sal, p_mean):
 
     ax.set_xticks(range(n_nets))
     ax.set_yticks(range(n_nets))
-    ax.set_xticklabels(_NET_NAMES, rotation=40, ha="right", fontsize=9, color="#ccc")
-    ax.set_yticklabels(_NET_NAMES, fontsize=9, color="#ccc")
+    ax.set_xticklabels(_nn, rotation=40, ha="right", fontsize=9, color="#ccc")
+    ax.set_yticklabels(_nn, fontsize=9, color="#ccc")
     ax.tick_params(colors="#555", length=0)
     for sp in ax.spines.values():
         sp.set_color("#222")
@@ -173,7 +222,7 @@ def _saliency_figure(sal, p_mean):
 
     # Callout labels for top-3 cross-network edges
     for rank, (score, i, j) in enumerate(top3_edges):
-        label = f"#{rank+1} {_NET_NAMES[i]}↔{_NET_NAMES[j]}"
+        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=6, color="#fb923c", fontweight="600",
@@ -193,9 +242,9 @@ def _saliency_figure(sal, p_mean):
 
     order = net_imp.argsort()[::-1]
     bars  = ax2.barh(range(n_nets), net_imp[order],
-                     color=[_NET_COLORS[i] for i in order], alpha=0.88, edgecolor="none", height=0.65)
+                     color=[_nc[i] for i in order], alpha=0.88, edgecolor="none", height=0.65)
     ax2.set_yticks(range(n_nets))
-    ax2.set_yticklabels([_NET_NAMES[i] for i in order], fontsize=9.5, color="#ddd")
+    ax2.set_yticklabels([_nn[i] for i in order], fontsize=9.5, color="#ddd")
     ax2.set_xlabel("Mean gradient magnitude", color="#555", fontsize=9)
     ax2.set_title("Network Importance for This Prediction", color="#bbb", fontsize=11, pad=14, fontweight="bold")
     ax2.invert_yaxis()
@@ -250,7 +299,7 @@ def _saliency_figure(sal, p_mean):
     ax3.set_box_aspect([1.2, 1.4, 1.0])
 
     fig.suptitle(
-        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  {len(_models)}-model LOSO ensemble  ·  CC200 → Yeo-7 networks",
+        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  {len(models)}-model LOSO ensemble  ·  CC200 → Yeo-7 networks",
         color="#444", fontsize=8.5, y=1.02,
     )
     plt.tight_layout()
@@ -267,6 +316,7 @@ def run_gcn(file_path):
         return "", "", "", None
 
     path = Path(file_path)
+    atlas_key = "cc200"  # default; overridden below for .1D files
     try:
         if path.suffix == ".npz":
             d   = np.load(path, allow_pickle=True)
@@ -282,13 +332,40 @@ def run_gcn(file_path):
             adj_t = torch.FloatTensor(adj).unsqueeze(0)
         else:
             bold = np.loadtxt(path, dtype=np.float32)
-            if bold.ndim != 2 or bold.shape[1] != 200:
-                return f"Error: expected (T×200), got {bold.shape}", "", "", None
+            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
 
-    models = get_models()
+
+
+    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:
@@ -300,7 +377,12 @@ def run_gcn(file_path):
     conf      = max(p_mean, 1 - p_mean) * 100
 
     try:
-        sal_img = _saliency_figure(_compute_saliency(bw_t, adj_t, models), p_mean)
+        sal_img = _saliency_figure(
+            _compute_saliency(bw_t, adj_t, models), p_mean,
+            net_names=atlas_cfg["net_names"],
+            net_bounds=atlas_cfg["net_bounds"],
+            net_colors=atlas_cfg["net_colors"],
+        )
     except Exception:
         sal_img = None
 
@@ -398,7 +480,7 @@ LOSO AUC = 0.7872 · 529 held-out subjects · 4 institutions
   <div><div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:3px">ICD-10 Classification</div>
   <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.4">{icd}</div></div>
   <div><div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1px;margin-bottom:3px">Ensemble Confidence</div>
-  <div style="color:#cbd5e1;font-size:0.84rem">{conf:.1f}% · p(ASD) = {p_mean:.3f} · {len(_models)}-model LOSO</div></div>
+  <div style="color:#cbd5e1;font-size:0.84rem">{conf:.1f}% · p(ASD) = {p_mean:.3f} · {len(models)}-model LOSO</div></div>
 </div>
 
 <div style="color:#8b95a7;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:4px;font-weight:500">Impression</div>
@@ -576,7 +658,7 @@ ARCHITECTURE = """
     <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 × 200 ROIs</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>
@@ -637,7 +719,7 @@ ARCHITECTURE = """
   <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 · 17 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 (Craddock 2012) · 200 functional 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">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=16 modes · hidden_dim=64</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">LOSO AUC = <span style="color:#ef4444;font-weight:600">0.7872</span> · 529 held-out subjects · 0 confident misclassifications</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>
@@ -715,7 +797,7 @@ with gr.Blocks(title="BrainConnect-ASD", css=css, theme=gr.themes.Base()) as dem
 
     with gr.Tabs():
         with gr.Tab("Analysis"):
-            file_input = gr.File(label="Upload CC200 fMRI (.1D or .npz)", type="filepath")
+            file_input = gr.File(label="Upload fMRI time series — CC200 (200), AAL (116), or Harvard-Oxford (111) ROIs · .1D or .npz", type="filepath")
             gr.HTML("<div style='color:#8b95a7;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.2px;margin:14px 0 8px;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")

app_with_llm.py

@@ -0,0 +1,363 @@
+"""
+BrainConnect-ASD — Scanner-site-invariant ASD detection from fMRI.
+Full pipeline: Adversarial GCN + Qwen2.5-7B fine-tuned on AMD MI300X.
+"""
+from __future__ import annotations
+
+import io
+from pathlib import Path
+
+import numpy as np
+import torch
+import gradio as gr
+
+_WINDOW_LEN   = 50
+_STEP         = 3
+_MAX_WINDOWS  = 30
+_FC_THRESHOLD = 0.2
+
+_CKPTS = {
+    "NYU":  Path("checkpoints/nyu.ckpt"),
+    "USM":  Path("checkpoints/usm.ckpt"),
+    "UCLA": Path("checkpoints/ucla.ckpt"),
+    "UM":   Path("checkpoints/um.ckpt"),
+}
+
+_LLM_MODEL = "Yatsuiii/asd-interpreter-lora"
+
+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 and provide structured clinical summaries for neurologists "
+    "and psychiatrists. Your reports are informative and evidence-based but always "
+    "clarify that findings are AI-assisted and should be integrated with full "
+    "clinical assessment. You do not make a diagnosis."
+)
+
+# ── 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)
+
+# ── GCN model loading ──────────────────────────────────────────────────────
+
+_models: list | None = None
+
+def get_models():
+    global _models
+    if _models is not None:
+        return _models
+    from brain_gcn.tasks import ClassificationTask
+    _models = []
+    for site, ckpt in _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))
+    return _models
+
+# ── LLM loading ────────────────────────────────────────────────────────────
+
+_llm = None
+
+def get_llm():
+    global _llm
+    if _llm is not None:
+        return _llm
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    print(f"Loading LLM: {_LLM_MODEL}")
+    tok = AutoTokenizer.from_pretrained(_LLM_MODEL)
+    tok.pad_token = tok.eos_token
+    mdl = AutoModelForCausalLM.from_pretrained(
+        _LLM_MODEL,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+    )
+    mdl.eval()
+    _llm = (mdl, tok)
+    return _llm
+
+
+def _llm_report(p_mean: float, per_model: list) -> 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)"
+    )
+    user_msg = (
+        f"Brain Connectivity GCN Analysis Report\n"
+        f"{'='*40}\n"
+        f"p(ASD)           : {p_mean:.3f}\n"
+        f"Confidence Level : {conf_label}\n"
+        f"Model Consensus  : {consensus}/4 site-blind models predict ASD\n\n"
+        f"Per-Model Breakdown (LOSO ensemble):\n{per_model_str}\n\n"
+        f"Please provide a structured clinical interpretation of these findings."
+    )
+    try:
+        mdl, tok = get_llm()
+        messages = [
+            {"role": "system", "content": SYSTEM_PROMPT},
+            {"role": "user",   "content": user_msg},
+        ]
+        text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = tok(text, return_tensors="pt").to(next(mdl.parameters()).device)
+        with torch.no_grad():
+            out = mdl.generate(
+                **inputs,
+                max_new_tokens=512,
+                temperature=0.3,
+                do_sample=True,
+                pad_token_id=tok.eos_token_id,
+            )
+        generated = out[0][inputs["input_ids"].shape[1]:]
+        return tok.decode(generated, skip_special_tokens=True).strip()
+    except Exception as e:
+        return f"LLM unavailable: {e}"
+
+# ── gradient saliency ──────────────────────────────────────────────────────
+
+def _compute_saliency(bw_t: torch.Tensor, adj_t: torch.Tensor, models) -> np.ndarray:
+    maps = []
+    for _, task in models:
+        adj = adj_t.clone().requires_grad_(True)
+        logits = task.model(bw_t, adj)
+        p = torch.softmax(logits, -1)[0, 1]
+        p.backward()
+        maps.append(adj.grad[0].abs().detach().numpy())
+    sal = np.mean(maps, axis=0)
+    sal = (sal + sal.T) / 2
+    return sal
+
+
+def _saliency_figure(sal: np.ndarray, p_mean: float):
+    import matplotlib
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
+    from PIL import Image
+
+    thresh  = np.percentile(sal, 95)
+    sal_top = np.where(sal >= thresh, sal, 0.0)
+    roi_imp = sal.sum(1)
+    top20   = roi_imp.argsort()[-20:][::-1]
+    verdict_color = "#e63946" if p_mean > 0.6 else "#2dc653" if p_mean < 0.4 else "#f4a261"
+
+    fig, axes = plt.subplots(1, 2, figsize=(14, 5.5))
+    fig.patch.set_facecolor("#0d0d0d")
+
+    ax = axes[0]
+    ax.set_facecolor("#111")
+    im = ax.imshow(sal_top, cmap="inferno", aspect="auto", interpolation="nearest")
+    ax.set_title("FC Edge Saliency  (top 5% connections)", color="#ccc", fontsize=11, pad=10)
+    ax.set_xlabel("ROI index", color="#777", fontsize=9)
+    ax.set_ylabel("ROI index", color="#777", fontsize=9)
+    ax.tick_params(colors="#555", labelsize=8)
+    cb = plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
+    cb.ax.yaxis.set_tick_params(color="#555", labelsize=7)
+    plt.setp(cb.ax.yaxis.get_ticklabels(), color="#666")
+    for spine in ax.spines.values():
+        spine.set_color("#333")
+
+    ax2 = axes[1]
+    ax2.set_facecolor("#111")
+    ax2.barh(range(20), roi_imp[top20], color=verdict_color, alpha=0.75, edgecolor="none")
+    ax2.set_yticks(range(20))
+    ax2.set_yticklabels([f"ROI {i:03d}" for i in top20], fontsize=8, color="#ccc")
+    ax2.set_xlabel("Cumulative gradient magnitude", color="#777", fontsize=9)
+    ax2.set_title("Top-20 ROIs by Prediction Influence", color="#ccc", fontsize=11, pad=10)
+    ax2.tick_params(colors="#555", labelsize=8)
+    ax2.invert_yaxis()
+    for spine in ["top", "right"]:
+        ax2.spines[spine].set_visible(False)
+    for spine in ["bottom", "left"]:
+        ax2.spines[spine].set_color("#333")
+
+    fig.suptitle(
+        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  Ensemble of {len(_models)} LOSO models",
+        color="#888", fontsize=10, y=1.02,
+    )
+    plt.tight_layout()
+    buf = io.BytesIO()
+    plt.savefig(buf, format="png", dpi=120, bbox_inches="tight", facecolor="#0d0d0d")
+    plt.close(fig)
+    buf.seek(0)
+    return Image.open(buf).copy()
+
+# ── inference ──────────────────────────────────────────────────────────────
+
+def run_gcn(file_path: str | None):
+    if file_path is None:
+        return "", "", "", None, ""
+
+    path = Path(file_path)
+    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 or bold.shape[1] != 200:
+                return f"⚠️ Error: expected (T×200) array, got {bold.shape}", "", "", None, ""
+            bw_t, adj_t = preprocess(bold)
+    except Exception as e:
+        return f"⚠️ Error loading file: {e}", "", "", None, ""
+
+    models = get_models()
+
+    per_model = []
+    with torch.no_grad():
+        for site, task in models:
+            logits = task(bw_t, adj_t)
+            p = torch.softmax(logits, -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
+
+    try:
+        sal = _compute_saliency(bw_t, adj_t, models)
+        sal_img = _saliency_figure(sal, p_mean)
+    except Exception:
+        sal_img = None
+
+    # Verdict
+    if p_mean > 0.6:
+        verdict = f"""<div style="background:#1a1a2e;border-left:6px solid #e63946;padding:24px 28px;border-radius:12px;margin-bottom:8px">
+<div style="font-size:2rem;font-weight:800;color:#e63946;letter-spacing:1px">ASD INDICATED</div>
+<div style="font-size:1.1rem;color:#aaa;margin-top:6px">Confidence: <b style="color:white">{conf:.1f}%</b> &nbsp;|&nbsp; p(ASD) = <b style="color:white">{p_mean:.3f}</b> &nbsp;|&nbsp; <b style="color:white">{consensus}/4</b> site-blind models agree</div>
+</div>"""
+    elif p_mean < 0.4:
+        verdict = f"""<div style="background:#1a1a2e;border-left:6px solid #2dc653;padding:24px 28px;border-radius:12px;margin-bottom:8px">
+<div style="font-size:2rem;font-weight:800;color:#2dc653;letter-spacing:1px">TYPICAL CONTROL</div>
+<div style="font-size:1.1rem;color:#aaa;margin-top:6px">Confidence: <b style="color:white">{conf:.1f}%</b> &nbsp;|&nbsp; p(ASD) = <b style="color:white">{p_mean:.3f}</b> &nbsp;|&nbsp; <b style="color:white">{4-consensus}/4</b> site-blind models agree</div>
+</div>"""
+    else:
+        verdict = f"""<div style="background:#1a1a2e;border-left:6px solid #f4a261;padding:24px 28px;border-radius:12px;margin-bottom:8px">
+<div style="font-size:2rem;font-weight:800;color:#f4a261;letter-spacing:1px">INCONCLUSIVE</div>
+<div style="font-size:1.1rem;color:#aaa;margin-top:6px">Confidence: <b style="color:white">{conf:.1f}%</b> &nbsp;|&nbsp; p(ASD) = <b style="color:white">{p_mean:.3f}</b> &nbsp;|&nbsp; Model disagreement — clinical review required</div>
+</div>"""
+
+    # Ensemble breakdown
+    rows = ""
+    for site, p in per_model:
+        lbl   = "ASD" if p > 0.5 else "TC"
+        color = "#e63946" if p > 0.5 else "#2dc653"
+        bar_w = int(p * 100)
+        rows += f"""<tr>
+<td style="padding:8px 12px;color:#ccc;font-weight:600">{site}-blind</td>
+<td style="padding:8px 12px"><div style="background:#333;border-radius:4px;height:18px;width:160px">
+<div style="background:{color};height:18px;width:{bar_w}%;border-radius:4px;opacity:0.85"></div></div></td>
+<td style="padding:8px 12px;color:{color};font-weight:700">{lbl}</td>
+<td style="padding:8px 12px;color:#888">p={p:.3f}</td>
+</tr>"""
+
+    ensemble = f"""<div style="background:#111;border-radius:10px;padding:20px;margin-top:4px">
+<div style="color:#888;font-size:0.8rem;text-transform:uppercase;letter-spacing:2px;margin-bottom:14px">Leave-One-Site-Out Ensemble — each model never trained on that site's data</div>
+<table style="width:100%;border-collapse:collapse">{rows}</table>
+<div style="margin-top:14px;color:#666;font-size:0.82rem">Cross-site consensus: {consensus}/4 models agree &nbsp;·&nbsp; LOSO AUC = 0.7872 across 529 held-out subjects</div>
+</div>"""
+
+    # LLM clinical report
+    llm_text = _llm_report(p_mean, per_model)
+    report = f"""<div style="background:#111;border-radius:10px;padding:20px;margin-top:4px">
+<div style="color:#888;font-size:0.8rem;text-transform:uppercase;letter-spacing:2px;margin-bottom:14px">Clinical Report — Qwen2.5-7B fine-tuned on AMD Instinct MI300X</div>
+<div style="color:#ddd;font-size:0.95rem;line-height:1.7;white-space:pre-wrap">{llm_text}</div>
+<div style="background:#1a1a1a;border-radius:6px;padding:12px;color:#555;font-size:0.78rem;margin-top:16px">
+⚕️ AI-assisted analysis only. Does not constitute a diagnosis. Integrate with clinical history, behavioral assessment, and standardized instruments (ADOS-2, ADI-R).
+</div></div>"""
+
+    return verdict, ensemble, report, sal_img
+
+
+# ── UI ─────────────────────────────────────────────────────────────────────
+
+css = """
+body { background: #0d0d0d; }
+.gradio-container { max-width: 960px; margin: auto; }
+"""
+
+with gr.Blocks(title="BrainConnect-ASD", css=css, theme=gr.themes.Base()) as demo:
+    gr.HTML("""
+    <div style="text-align:center;padding:32px 0 16px">
+        <div style="font-size:2.2rem;font-weight:900;color:white;letter-spacing:-1px">BrainConnect<span style="color:#e63946">-ASD</span></div>
+        <div style="color:#888;font-size:1rem;margin-top:8px">Scanner-site-invariant ASD detection from resting-state fMRI</div>
+        <div style="display:flex;justify-content:center;gap:24px;margin-top:16px;flex-wrap:wrap">
+            <span style="background:#1a1a2e;color:#aaa;padding:6px 14px;border-radius:20px;font-size:0.85rem">LOSO AUC 0.7872</span>
+            <span style="background:#1a1a2e;color:#aaa;padding:6px 14px;border-radius:20px;font-size:0.85rem">529 held-out subjects</span>
+            <span style="background:#1a1a2e;color:#aaa;padding:6px 14px;border-radius:20px;font-size:0.85rem">4 independent institutions</span>
+            <span style="background:#1a1a2e;color:#aaa;padding:6px 14px;border-radius:20px;font-size:0.85rem">AMD Instinct MI300X</span>
+        </div>
+    </div>
+    """)
+
+    file_input = gr.File(label="Upload CC200 fMRI file (.1D or .npz)", type="filepath")
+    verdict_html  = gr.HTML()
+    ensemble_html = gr.HTML()
+
+    gr.HTML("<div style='color:#888;font-size:0.8rem;text-transform:uppercase;letter-spacing:2px;margin:24px 0 8px'>Gradient Saliency — which brain connections drove this prediction</div>")
+    saliency_img = gr.Image(label="FC Edge Saliency & ROI Importance", type="pil")
+
+    report_html = gr.HTML()
+
+    file_input.change(
+        fn=run_gcn,
+        inputs=file_input,
+        outputs=[verdict_html, ensemble_html, report_html, saliency_img],
+    )
+
+    gr.HTML("""
+    <div style="text-align:center;padding:24px 0;color:#444;font-size:0.8rem">
+        Adversarial Brain-Mode GCN (k=16) · Qwen2.5-7B LoRA (AMD MI300X) · ABIDE I ·
+        <a href="https://github.com/Yatsuiii/Brain-Connectivity-GCN" style="color:#666">GitHub</a>
+    </div>
+    """)
+
+print("Preloading GCN models...")
+get_models()
+print("Preloading LLM...")
+get_llm()
+print("All models ready.")
+
+if __name__ == "__main__":
+    demo.launch()

checkpoints/aal_nyu.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d588597092f4b9483fdcfb0f8d5aae67030ea9d17df8cf7a0c027ef527b5657
+size 253386

checkpoints/aal_ucla.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3af986f0e5909e8551c2473ee3ff876ff9cc814ea38e220de33fad96201eaa37
+size 253386

checkpoints/aal_um.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b380df053b6a9a83dc6062070418cd6b8718b4f4c5c5f8b388330dcf0a9abf5
+size 253386

checkpoints/aal_usm.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7410692b93237e59fe5a03726338b7f02f0b61acbb6185dc1ecc79dfbee6bdac
+size 252813