Add EN/FR bilingual UI + Hôpital Montfort theme (Montfort blue + turquoise)

app.py

@@ -1,4 +1,4 @@
-"""HF Spaces Gradio app — chronic wound classifier (4-class).
+"""HF Spaces Gradio app — chronic wound classifier (4-class), bilingual EN/FR.
 
 Self-contained: no wound_classifier package install required. The model
 architecture and transforms are inlined here so this file plus the .pt
@@ -7,6 +7,10 @@ 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.
+
+Theming approximates Hôpital Montfort (Ottawa) brand colors, sourced from
+the live hopitalmontfort.com stylesheet: primary "Montfort blue" #00729a,
+turquoise accent #47c9cd, warm cream surface #f1ede5.
 """
 
 from __future__ import annotations
@@ -24,14 +28,166 @@ 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"]
 
+CLASS_NAMES: dict[str, dict[str, str]] = {
+    "en": {
+        "D": "Diabetic ulcer",
+        "P": "Pressure ulcer",
+        "S": "Surgical wound",
+        "V": "Venous ulcer",
+    },
+    "fr": {
+        "D": "Ulcère diabétique",
+        "P": "Escarre",
+        "S": "Plaie chirurgicale",
+        "V": "Ulcère veineux",
+    },
+}
+
+LOW_CONFIDENCE_THRESHOLD = 0.5
+
+# ---------- Localized strings ---------------------------------------------------
+
+TITLE: dict[str, str] = {
+    "en": "# Chronic Wound Classifier — 4-class AZH demo",
+    "fr": "# Classification des plaies chroniques — démo AZH 4 classes",
+}
+
+DESCRIPTION: dict[str, str] = {
+    "en": """\
+Upload a close-up photo of a chronic wound and the model returns its best guess at the wound type
+(diabetic ulcer, pressure ulcer, surgical wound, or venous ulcer) with per-class probabilities.
+
+**Before you upload — a few things worth knowing:**
+
+- **Frame the wound in the center.** The model resizes and center-crops to 224×224 pixels, so a wound
+  off in the corner of a high-resolution phone photo will get cropped out. Close-up clinical-style
+  framing (the wound fills most of the frame) gives the cleanest predictions.
+- **JPEG or PNG works fine.** Anything PIL can open.
+- **There is no "not a wound" option.** Upload a face, a sandwich, a sunset — the model will
+  confidently place it into one of the four wound classes anyway. The low-confidence warning below
+  is the only signal you'll get that the image probably isn't what the model was trained on.
+- **Pressure-ulcer predictions are weakest** (~41% accuracy on the held-out test set). When the model
+  says "Pressure ulcer," treat the call with extra skepticism.
+
+**This is a research demo, not a medical device.** It does not diagnose, triage, or replace clinician
+judgement. 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).
+""",
+    "fr": """\
+Téléversez une photo en gros plan d'une plaie chronique : le modèle propose son meilleur diagnostic
+parmi quatre types de plaies (ulcère diabétique, escarre, plaie chirurgicale, ulcère veineux) avec
+les probabilités par classe.
+
+**À lire avant de téléverser :**
+
+- **Cadrez la plaie au centre de l'image.** Le modèle redimensionne et recadre au centre à 224×224 pixels :
+  une plaie située dans un coin d'une photo haute résolution sera coupée. Un cadrage clinique rapproché
+  (la plaie occupe la majeure partie du cadre) donne les prédictions les plus propres.
+- **JPEG ou PNG conviennent.** Tout format que PIL sait ouvrir.
+- **Il n'y a pas d'option « ce n'est pas une plaie ».** Téléversez un visage, un sandwich, un coucher
+  de soleil : le modèle classera l'image dans l'une des quatre classes de plaies, et avec confiance.
+  L'avertissement de faible confiance ci-dessous est le seul signal indiquant que l'image n'est
+  probablement pas conforme aux données d'entraînement.
+- **Les prédictions d'escarre sont les moins fiables** (~41 % d'exactitude sur le jeu de test).
+  Lorsque le modèle indique « Escarre », à interpréter avec une prudence accrue.
+
+**Ceci est une démonstration de recherche, et non un dispositif médical.** Le modèle n'établit aucun
+diagnostic, ne fait pas de triage et ne remplace en rien le jugement clinique. Entraîné sur la base
+de données publique AZH Chronic Wound Database (Anisuzzaman et coll. 2022) avec EfficientNet-B0 et
+validation croisée à 10 plis groupés par patient ; le point de contrôle déployé
+(`cv_baseline_fold5_best.pt`) atteint **0,8152 top-1** sur le jeu de test AZH retenu (n=184).
+""",
+}
+
+ARTICLE: dict[str, str] = {
+    "en": """\
+**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). No mobile / offline build.
+
+**Citation.** Anisuzzaman et al. 2022, *Multi-modal wound classification using wound image and
+location by deep neural network*, Sci. Rep. 12:20057.
+""",
+    "fr": """\
+**Méthodologie en un paragraphe.** EfficientNet-B0 (préentraîné sur ImageNet), affinement en deux
+phases : la phase 1 gèle le tronc et entraîne la tête à 4 classes pendant 5 époques (lr=1e-3) ; la
+phase 2 dégèle et affine le réseau complet pendant 15 époques (lr=1e-4). La validation croisée à
+10 plis groupés par patient garantit que les images d'un même patient n'apparaissent jamais à la
+fois en entraînement et en validation. Les découpes aléatoires stratifiées (sans regroupement par
+patient) gonflent l'exactitude de ~3 à 30 pp sur AZH (mesuré directement) ; nous rapportons les
+chiffres groupés par patient comme évaluation honnête. Le point de contrôle déployé est la meilleure
+fold individuelle sur le jeu de test retenu ; un ensemble par vote doux à 10 plis atteint en moyenne
+0,7989 sur ce même jeu.
+
+**Hors champ.** Aucun usage de décision clinique. Aucune prétention à une exactitude diagnostique
+sur de vraies cohortes de patients. Aucun audit d'équité par teinte de peau (limite connue). Aucune
+version mobile ou hors ligne.
+
+**Référence.** Anisuzzaman et coll. 2022, *Multi-modal wound classification using wound image and
+location by deep neural network*, Sci. Rep. 12:20057.
+""",
+}
+
+LABELS: dict[str, dict[str, str]] = {
+    "en": {
+        "lang_radio": "Language / Langue",
+        "image_input": "Wound photograph (close-up, centered)",
+        "label_output": "Predicted wound type",
+        "notes_output": "Notes",
+        "submit": "Classify",
+        "clear": "Clear",
+    },
+    "fr": {
+        "lang_radio": "Language / Langue",
+        "image_input": "Photographie de la plaie (gros plan, centrée)",
+        "label_output": "Type de plaie prédit",
+        "notes_output": "Remarques",
+        "submit": "Classer",
+        "clear": "Effacer",
+    },
+}
+
+NOTE_LOW_CONFIDENCE: dict[str, str] = {
+    "en": (
+        "⚠️ **Low confidence** (top class {top_label} at {top_pct}). "
+        "This often means the photo isn't a clear close-up of a wound, or the wound type "
+        'isn\'t one of the four the model was trained on. The model has no "not a wound" '
+        "option — it will always pick one of D / P / S / V even when the image isn't a wound."
+    ),
+    "fr": (
+        "⚠️ **Faible confiance** (classe principale {top_label} à {top_pct}). "
+        "Cela indique souvent que la photo n'est pas un gros plan clair d'une plaie, ou "
+        "que le type de plaie ne fait pas partie des quatre sur lesquels le modèle a été entraîné. "
+        "Le modèle n'a pas d'option « ce n'est pas une plaie » — il choisira toujours D / P / S / V "
+        "même si l'image n'est pas une plaie."
+    ),
+}
+
+NOTE_PRESSURE: dict[str, str] = {
+    "en": (
+        "⚠️ **Pressure-class predictions are the model's weak spot** "
+        "(~0.41 accuracy on the held-out test set). Treat this prediction with extra skepticism."
+    ),
+    "fr": (
+        "⚠️ **Les prédictions d'escarre sont le point faible du modèle** "
+        "(~0,41 d'exactitude sur le jeu de test retenu). Interpréter cette prédiction avec une "
+        "prudence accrue."
+    ),
+}
+
+
+# ---------- Model loading -------------------------------------------------------
+
 
 def _build_model(num_classes: int = 4) -> nn.Module:
     model: nn.Module = efficientnet_b0(weights=None)
@@ -66,87 +222,179 @@ MODEL = _load_model(CKPT_PATH)
 TRANSFORM = _build_transform()
 
 
-LOW_CONFIDENCE_THRESHOLD = 0.5
+def _lang_code(choice: str) -> str:
+    return "fr" if choice == "Français" else "en"
 
 
-def classify(image: Image.Image) -> tuple[dict[str, float], str]:
+def _format_pct(value: float, lang: str) -> str:
+    pct = f"{value:.0%}"
+    # Use French non-breaking space + lowercase percent? Standard French formatting.
+    return pct.replace(".", ",") if lang == "fr" else pct
+
+
+def classify(image: Image.Image | None, language_choice: str) -> tuple[dict[str, float], str]:
     if image is None:
         return {}, ""
+    lang = _lang_code(language_choice)
     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()
-    label_probs = {CLASS_NAMES[IDX_TO_CLASS[i]]: float(probs[i]) for i in range(4)}
+
+    name_map = CLASS_NAMES[lang]
+    label_probs = {name_map[IDX_TO_CLASS[i]]: float(probs[i]) for i in range(4)}
 
     top_label, top_prob = max(label_probs.items(), key=lambda kv: kv[1])
+    top_letter = next(letter for letter, name in name_map.items() if name == top_label)
+
     notes: list[str] = []
     if top_prob < LOW_CONFIDENCE_THRESHOLD:
         notes.append(
-            f"⚠️ **Low confidence** (top class {top_label} at {top_prob:.0%}). "
-            "This often means the photo isn't a clear close-up of a wound, or the wound type "
-            "isn't one of the four the model was trained on. The model has no \"not a wound\" "
-            "option — it will always pick one of D / P / S / V even when the image isn't a wound."
-        )
-    if top_label == "Pressure ulcer":
-        notes.append(
-            "⚠️ **Pressure-class predictions are the model's weak spot** "
-            "(~0.41 accuracy on the held-out test set). Treat this prediction with extra skepticism."
+            NOTE_LOW_CONFIDENCE[lang].format(
+                top_label=top_label, top_pct=_format_pct(top_prob, lang)
+            )
         )
+    if top_letter == "P":
+        notes.append(NOTE_PRESSURE[lang])
+
     return label_probs, "\n\n".join(notes)
 
 
-DESCRIPTION = """\
-Upload a close-up photo of a chronic wound and the model returns its best guess at the wound type
-(diabetic ulcer, pressure ulcer, surgical wound, or venous ulcer) with per-class probabilities.
+# ---------- UI ------------------------------------------------------------------
 
-**Before you upload — a few things worth knowing:**
+# Custom theme using Hôpital Montfort brand colors (extracted from their stylesheet):
+#   primary "Montfort blue" #00729a, turquoise accent #47c9cd, warm cream #f1ede5.
+montfort_blue = gr.themes.Color(
+    name="montfort_blue",
+    c50="#eef7fa",
+    c100="#c6eafa",
+    c200="#9bd9ed",
+    c300="#6ec5dd",
+    c400="#3aa5c4",
+    c500="#00729a",
+    c600="#005f81",
+    c700="#004d68",
+    c800="#003a4f",
+    c900="#002836",
+    c950="#001a25",
+)
+montfort_turquoise = gr.themes.Color(
+    name="montfort_turquoise",
+    c50="#e6fbfb",
+    c100="#c6f4f5",
+    c200="#9eeaeb",
+    c300="#73dde0",
+    c400="#47c9cd",
+    c500="#23b6ba",
+    c600="#1a9498",
+    c700="#147576",
+    c800="#0f5859",
+    c900="#0a3c3d",
+    c950="#062323",
+)
 
-- **Frame the wound in the center.** The model resizes and center-crops to 224×224 pixels, so a wound
-  off in the corner of a high-resolution phone photo will get cropped out. Close-up clinical-style
-  framing (the wound fills most of the frame) gives the cleanest predictions.
-- **JPEG or PNG works fine.** Anything PIL can open.
-- **There is no "not a wound" option.** Upload a face, a sandwich, a sunset — the model will
-  confidently place it into one of the four wound classes anyway. The low-confidence warning below
-  is the only signal you'll get that the image probably isn't what the model was trained on.
-- **Pressure-ulcer predictions are weakest** (~41% accuracy on the held-out test set). When the model
-  says "Pressure ulcer," treat the call with extra skepticism.
+theme = gr.themes.Soft(
+    primary_hue=montfort_blue,
+    secondary_hue=montfort_turquoise,
+    neutral_hue=gr.themes.colors.stone,
+    font=[gr.themes.GoogleFont("Inter"), "system-ui", "sans-serif"],
+).set(
+    body_background_fill="#f1ede5",
+    block_background_fill="white",
+    button_primary_background_fill="#00729a",
+    button_primary_background_fill_hover="#005f81",
+    button_primary_text_color="white",
+)
 
-**This is a research demo, not a medical device.** It does not diagnose, triage, or replace clinician
-judgement. 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).
-"""
 
-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.
+def _localize_components(
+    language_choice: str,
+) -> tuple[gr.Markdown, gr.Markdown, gr.Image, gr.Label, gr.Markdown, gr.Button, gr.Button]:
+    lang = _lang_code(language_choice)
+    labels = LABELS[lang]
+    return (
+        gr.Markdown(value=TITLE[lang]),
+        gr.Markdown(value=DESCRIPTION[lang]),
+        gr.Image(label=labels["image_input"]),
+        gr.Label(label=labels["label_output"]),
+        gr.Markdown(value="", label=labels["notes_output"]),
+        gr.Button(value=labels["submit"]),
+        gr.Button(value=labels["clear"]),
+    )
 
-**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.
-"""
+with gr.Blocks(theme=theme, title="Chronic Wound Classifier · Hôpital Montfort demo") as demo:
+    language_radio = gr.Radio(
+        choices=["English", "Français"],
+        value="English",
+        label=LABELS["en"]["lang_radio"],
+        interactive=True,
+    )
 
+    title_md = gr.Markdown(TITLE["en"])
+    description_md = gr.Markdown(DESCRIPTION["en"])
 
-demo = gr.Interface(
-    fn=classify,
-    inputs=gr.Image(type="pil", label="Wound photograph (close-up, centered)"),
-    outputs=[
-        gr.Label(num_top_classes=4, label="Predicted wound type"),
-        gr.Markdown(label="Notes"),
-    ],
-    title="Chronic Wound Classifier — 4-class AZH demo",
-    description=DESCRIPTION,
-    article=ARTICLE,
-    flagging_mode="never",
-)
+    with gr.Row():
+        with gr.Column():
+            image_input = gr.Image(type="pil", label=LABELS["en"]["image_input"])
+            with gr.Row():
+                submit_btn = gr.Button(LABELS["en"]["submit"], variant="primary")
+                clear_btn = gr.Button(LABELS["en"]["clear"])
+        with gr.Column():
+            label_output = gr.Label(num_top_classes=4, label=LABELS["en"]["label_output"])
+            notes_output = gr.Markdown(label=LABELS["en"]["notes_output"])
+
+    article_md = gr.Markdown(ARTICLE["en"])
+
+    submit_btn.click(
+        classify,
+        inputs=[image_input, language_radio],
+        outputs=[label_output, notes_output],
+    )
+    image_input.change(
+        classify,
+        inputs=[image_input, language_radio],
+        outputs=[label_output, notes_output],
+    )
+    clear_btn.click(
+        lambda: (None, {}, ""),
+        inputs=[],
+        outputs=[image_input, label_output, notes_output],
+    )
+
+    def _on_language_change(
+        language_choice: str, current_image: Image.Image | None
+    ) -> tuple[dict, dict, dict, dict, dict, dict, dict, str]:
+        lang = _lang_code(language_choice)
+        labels = LABELS[lang]
+        # Re-run inference so the on-screen probability labels switch languages too.
+        new_probs, new_notes = classify(current_image, language_choice)
+        return (
+            gr.update(value=TITLE[lang]),
+            gr.update(value=DESCRIPTION[lang]),
+            gr.update(value=ARTICLE[lang]),
+            gr.update(label=labels["image_input"]),
+            gr.update(label=labels["label_output"], value=new_probs),
+            gr.update(value=labels["submit"]),
+            gr.update(value=labels["clear"]),
+            new_notes,
+        )
+
+    language_radio.change(
+        _on_language_change,
+        inputs=[language_radio, image_input],
+        outputs=[
+            title_md,
+            description_md,
+            article_md,
+            image_input,
+            label_output,
+            submit_btn,
+            clear_btn,
+            notes_output,
+        ],
+    )
 
 
 if __name__ == "__main__":