Initial Space: Gradio app + cv_baseline_fold5 (Test top-1 0.8152)

README.md

@@ -1,13 +1,46 @@
 ---
-title: Wound Classifier
-emoji: 🐨
-colorFrom: gray
-colorTo: red
+title: Chronic Wound Classifier
+emoji: 🩹
+colorFrom: red
+colorTo: blue
 sdk: gradio
-sdk_version: 6.14.0
-python_version: '3.13'
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Chronic Wound Classifier — 4-class AZH demo
+
+Demo classifier for chronic wound photographs: predicts one of four wound types
+(diabetic ulcer, pressure ulcer, surgical wound, venous ulcer) from an uploaded image.
+
+**Not a medical device. Not for clinical use.** Research demonstration only.
+
+## Headline metric
+Top-1 accuracy on the held-out AZH Test set (n=184): **0.8152**
+(`cv_baseline_fold5_best.pt` — the highest single-fold checkpoint from
+patient-grouped 10-fold cross-validation).
+
+The 10-fold soft-vote ensemble of the same recipe scores 0.7989 on the same
+set; the single-checkpoint variant is shipped here for inference latency
+and footprint reasons.
+
+## Architecture
+EfficientNet-B0 (ImageNet-pretrained), two-phase fine-tune (head-only 5 epochs
+at lr=1e-3, then full network 15 epochs at lr=1e-4). Patient-grouped CV
+splits ensure the same patient's images never appear in both train and val.
+
+## Limitations
+- **Pressure-class accuracy is ~0.41** — interpret pressure-class predictions with care.
+- No fairness audit across skin tones (known gap).
+- English-only UI; no mobile or offline build.
+- Not validated on real patient cohorts outside AZH.
+
+## Source code & training pipeline
+The training, evaluation, and methodology code live in the project repo:
+[github.com — wound-classification](#) (full link to be added by user)
+
+## Citation
+Anisuzzaman et al. 2022. *Multi-modal wound classification using wound image
+and location by deep neural network.* Sci. Rep. 12:20057.

app.py

@@ -0,0 +1,123 @@
+"""HF Spaces Gradio app — chronic wound classifier (4-class).
+
+Self-contained: no wound_classifier package install required. The model
+architecture and transforms are inlined here so this file plus the .pt
+checkpoint and requirements.txt are everything the Space needs.
+
+If the architecture or transform here drifts from
+src/wound_classifier/{modeling/models.py, features.py} the Space and the
+training pipeline will silently disagree. Keep them in sync.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import gradio as gr
+import torch
+from PIL import Image
+from torch import nn
+from torchvision import transforms
+from torchvision.models import efficientnet_b0
+
+CKPT_PATH = Path(__file__).parent / "cv_baseline_fold5_best.pt"
+IMAGE_SIZE = 224
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+CLASS_NAMES = {
+    "D": "Diabetic ulcer",
+    "P": "Pressure ulcer",
+    "S": "Surgical wound",
+    "V": "Venous ulcer",
+}
+IDX_TO_CLASS = ["D", "P", "S", "V"]
+
+
+def _build_model(num_classes: int = 4) -> nn.Module:
+    model: nn.Module = efficientnet_b0(weights=None)
+    in_features = model.classifier[1].in_features  # type: ignore[index, union-attr]
+    model.classifier = nn.Sequential(
+        nn.Dropout(p=0.2, inplace=True),
+        nn.Linear(in_features, num_classes),
+    )
+    return model
+
+
+def _load_model(path: Path) -> nn.Module:
+    ckpt = torch.load(path, map_location="cpu", weights_only=False)
+    model = _build_model(num_classes=4)
+    model.load_state_dict(ckpt["state_dict"])
+    model.eval()
+    return model
+
+
+def _build_transform() -> transforms.Compose:
+    return transforms.Compose(
+        [
+            transforms.Resize(IMAGE_SIZE),
+            transforms.CenterCrop(IMAGE_SIZE),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
+        ]
+    )
+
+
+MODEL = _load_model(CKPT_PATH)
+TRANSFORM = _build_transform()
+
+
+def classify(image: Image.Image) -> dict[str, float]:
+    if image is None:
+        return {}
+    rgb = image.convert("RGB")
+    x = TRANSFORM(rgb).unsqueeze(0)
+    with torch.inference_mode():
+        logits = MODEL(x)
+        probs = torch.softmax(logits, dim=1).squeeze(0).numpy()
+    return {CLASS_NAMES[IDX_TO_CLASS[i]]: float(probs[i]) for i in range(4)}
+
+
+DESCRIPTION = """\
+Chronic wound photo classifier — predicts which of four wound types an uploaded image shows
+(diabetic ulcer, pressure ulcer, surgical wound, or venous ulcer).
+
+**This is a research demo, not a medical device.** It does not diagnose, triage, or replace clinician
+judgement. It was trained on the public AZH Chronic Wound Database (Anisuzzaman et al. 2022) using
+EfficientNet-B0 with patient-grouped 10-fold cross-validation; the deployed checkpoint
+(`cv_baseline_fold5_best.pt`) reaches **0.8152 top-1** on the held-out AZH Test set (n=184).
+Pressure ulcers are the underperforming class (~0.41 accuracy) — interpret pressure-class predictions
+with extra care.
+
+Upload a wound photograph below to see the model's per-class probabilities.
+"""
+
+ARTICLE = """\
+**Methodology in one paragraph.** EfficientNet-B0 (ImageNet-pretrained) two-phase fine-tuning:
+phase 1 freezes the backbone and trains the 4-class head for 5 epochs at lr=1e-3; phase 2 unfreezes
+and fine-tunes the full network for 15 epochs at lr=1e-4. Patient-grouped 10-fold CV ensures the same
+patient's images never appear in both train and val. Training and validation under random-stratified
+splits would inflate accuracy by ~3–30pp on AZH (we measured this directly); we report the patient-
+grouped numbers as the honest evaluation. The shipped checkpoint is the highest single fold on the
+held-out Test set; a 10-fold soft-vote ensemble averages 0.7989 on the same set.
+
+**Out of scope.** Not for clinical decision-making. No claim of diagnostic accuracy on real patient
+cohorts. No fairness audit across skin tones (known gap). English-only UI. No mobile / offline build.
+
+**Citations.** Anisuzzaman et al. 2022, *Multi-modal wound classification using wound image and
+location by deep neural network*, Sci. Rep. 12:20057.
+"""
+
+
+demo = gr.Interface(
+    fn=classify,
+    inputs=gr.Image(type="pil", label="Wound photograph"),
+    outputs=gr.Label(num_top_classes=4, label="Predicted wound type"),
+    title="Chronic Wound Classifier — 4-class AZH demo",
+    description=DESCRIPTION,
+    article=ARTICLE,
+    flagging_mode="never",
+)
+
+
+if __name__ == "__main__":
+    demo.launch()

cv_baseline_fold5_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f4b21f6204855827f33f2679be63d964d237a4d73a9940382ef4d49ed552e99
+size 16355126

requirements.txt

@@ -0,0 +1,4 @@
+torch>=2.0
+torchvision>=0.15
+gradio>=4.0
+pillow>=10.0