app.py

import gradio as gr
import numpy as np
from PIL import Image
import tensorflow as tf
from transformers import pipeline

# =========================
# 🔥 MODELO LOCAL
# =========================
model = tf.keras.models.load_model("modelo_derm_real.h5")

clases = ["carcinoma", "fibroma", "melanoma", "nevo", "precancer", "queratosis", "vascular"]

map_clinico = {
    "carcinoma": "Carcinoma basocelular",
    "fibroma": "Fibroma",
    "melanoma": "Melanoma",
    "nevo": "Nevo benigno",
    "precancer": "Lesión precancerosa",
    "queratosis": "Queratosis",
    "vascular": "Lesión vascular"
}

# =========================
# 🔥 MODELO LIGERO (fallback)
# =========================
hf_model = pipeline("image-classification", model="nateraw/vit-base-beans")

# =========================
# 🧠 ABCDE SIN OPENCV
# =========================
def analizar_abcde(image):
    img = np.array(image)

    # A: Asimetría
    h, w, _ = img.shape
    izquierda = img[:, :w//2]
    derecha = np.fliplr(img[:, w//2:])
    min_w = min(izquierda.shape[1], derecha.shape[1])
    asimetria = np.mean(np.abs(izquierda[:, :min_w] - derecha[:, :min_w]))

    # B: Bordes (simple)
    bordes = np.std(img)

    # C: Color
    color_var = np.std(img)

    # D: Tamaño relativo
    diametro = max(h, w)

    score = 0

    if asimetria > 20:
        score += 1
    if bordes > 40:
        score += 1
    if color_var > 50:
        score += 1
    if diametro > 300:
        score += 1

    return score

# =========================
# 🔥 PREDICCIÓN
# =========================
def predict(image):
    try:
        image = image.convert("RGB")

        # ===== MODELO LOCAL =====
        img = image.resize((224, 224))
        arr = np.array(img)/255.0
        arr = np.expand_dims(arr, axis=0)

        pred = model.predict(arr, verbose=0)[0]
        idx = np.argmax(pred)
        confianza = float(pred[idx]) * 100
        dx = clases[idx]
        dx_clinico = map_clinico[dx]

        # ===== ABCDE =====
        score_abcde = analizar_abcde(image)

        # ===== HF fallback =====
        hf = hf_model(image)
        hf_conf = hf[0]["score"] * 100

        # =========================
        # 🧠 LÓGICA CLÍNICA
        # =========================
        if score_abcde >= 3:
            diagnostico = "Sospechoso de malignidad"
            riesgo = "🔴 ALTO"
        elif score_abcde == 2:
            diagnostico = dx_clinico
            riesgo = "🟠 MODERADO"
        else:
            diagnostico = "Nevo benigno"
            riesgo = "🟢 BAJO"

        # 🔥 ANTI FALSOS POSITIVOS
        if dx_clinico in ["Melanoma", "Carcinoma basocelular"] and score_abcde < 3:
            diagnostico = "Nevo benigno"
            riesgo = "🟢 BAJO"

        return f"""
🏥 INFORME DERM-IA PRO

Diagnóstico: {diagnostico}
Confianza modelo: {round(confianza,2)}%
IA base: {round(hf_conf,2)}%

ABCDE score: {score_abcde}/4
Riesgo: {riesgo}

🩺 Conducta:
- 🔴 Biopsia urgente
- 🟠 Dermatoscopia
- 🟢 Seguimiento

⚠ Apoyo clínico.
"""

    except Exception as e:
        return f"❌ ERROR:\n{str(e)}"


# =========================
# 🎯 INTERFAZ
# =========================
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🧠 Derm-IA Clínico PRO")
    gr.Markdown("### IA dermatológica (estable + sin errores)")

    with gr.Row():
        img = gr.Image(type="pil")
        out = gr.Textbox(lines=15)

    btn = gr.Button("🔍 Analizar")
    btn.click(fn=predict, inputs=img, outputs=out)

demo.launch()