app.py

def analizar_lesion_combined(img):
    # Convertir imagen para Fastai
    img_fastai = PILImage.create(img)

    # ViT prediction
    inputs = feature_extractor(img, return_tensors="pt")
    with torch.no_grad():
        outputs = model_vit(**inputs)
        probs_vit = outputs.logits.softmax(dim=-1).cpu().numpy()[0]
    pred_idx_vit = int(np.argmax(probs_vit))
    pred_class_vit = CLASSES[pred_idx_vit]
    confidence_vit = probs_vit[pred_idx_vit]

    # Fast.ai models
    pred_fast_malignant, _, probs_fast_mal = model_malignancy.predict(img_fastai)
    prob_malignant = float(probs_fast_mal[1])  # índice 1 = maligno
    pred_fast_type, _, probs_fast_type = model_norm2000.predict(img_fastai)

    # Modelo TensorFlow ISIC (usando TFSMLayer)
    x_isic = preprocess_image_isic(img)
    preds_isic_dict = model_isic(x_isic)  # devuelve dict con tensores

    # DEBUG: imprime claves para saber cuál usar
    print("Claves de salida de model_isic:", preds_isic_dict.keys())

    # Por ejemplo, si la clave correcta es 'outputs', cámbiala aquí:
    # Si sólo hay una clave, la usamos directamente:
    key = list(preds_isic_dict.keys())[0]
    preds_isic = preds_isic_dict[key].numpy()[0]

    pred_idx_isic = int(np.argmax(preds_isic))
    pred_class_isic = CLASSES[pred_idx_isic]
    confidence_isic = preds_isic[pred_idx_isic]

    # Gráfico ViT
    colors_bars = [RISK_LEVELS[i]['color'] for i in range(7)]
    fig, ax = plt.subplots(figsize=(8, 3))
    ax.bar(CLASSES, probs_vit*100, color=colors_bars)
    ax.set_title("Probabilidad ViT por tipo de lesión")
    ax.set_ylabel("Probabilidad (%)")
    ax.set_xticks(np.arange(len(CLASSES)))
    ax.set_xticklabels(CLASSES, rotation=45, ha='right')
    ax.grid(axis='y', alpha=0.2)
    plt.tight_layout()
    buf = io.BytesIO()
    plt.savefig(buf, format="png")
    plt.close(fig)
    img_bytes = buf.getvalue()
    img_b64 = base64.b64encode(img_bytes).decode("utf-8")
    html_chart = f'<img src="data:image/png;base64,{img_b64}" style="max-width:100%"/>'

    # Informe HTML con los 4 modelos
    informe = f"""
    <div style="font-family:sans-serif; max-width:800px; margin:auto">
    <h2>🧪 Diagnóstico por 4 modelos de IA</h2>
    <table style="border-collapse: collapse; width:100%; font-size:16px">
        <tr><th style="text-align:left">🔍 Modelo</th><th>Resultado</th><th>Confianza</th></tr>
        <tr><td>🧠 ViT (transformer)</td><td><b>{pred_class_vit}</b></td><td>{confidence_vit:.1%}</td></tr>
        <tr><td>🧬 Fast.ai (clasificación)</td><td><b>{pred_fast_type}</b></td><td>N/A</td></tr>
        <tr><td>⚠️ Fast.ai (malignidad)</td><td><b>{"Maligno" if prob_malignant > 0.5 else "Benigno"}</b></td><td>{prob_malignant:.1%}</td></tr>
        <tr><td>🔬 ISIC TensorFlow</td><td><b>{pred_class_isic}</b></td><td>{confidence_isic:.1%}</td></tr>
    </table>
    <br>
    <b>🩺 Recomendación automática:</b><br> 
    """

    cancer_risk_score = sum(probs_vit[i] * RISK_LEVELS[i]['weight'] for i in range(7))
    if prob_malignant > 0.7 or cancer_risk_score > 0.6:
        informe += "🚨 <b>CRÍTICO</b> – Derivación urgente a oncología dermatológica"
    elif prob_malignant > 0.4 or cancer_risk_score > 0.4:
        informe += "⚠️ <b>ALTO RIESGO</b> – Consulta con dermatólogo en 7 días"
    elif cancer_risk_score > 0.2:
        informe += "📋 <b>RIESGO MODERADO</b> – Evaluación programada (2-4 semanas)"
    else:
        informe += "✅ <b>BAJO RIESGO</b> – Seguimiento de rutina (3-6 meses)"

    informe += "</div>"

    return informe, html_chart