app.py

# -*- coding: utf-8 -*-
"""
تحليل قرحة القدم باستخدام Unet + EfficientNet-b0
النموذج من Google Drive (best_model_5.pth)
"""

import os
import cv2
import gdown
import numpy as np
from PIL import Image
import torch
import gradio as gr
import segmentation_models_pytorch as smp

# =========================================================
# الإعدادات العامة
# =========================================================
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
IMG_SIZE = 512
THRESHOLD = 0.35
MODEL_PATH = "best_model_5.pth"
MODEL_URL = "https://drive.google.com/uc?id=1Ovaczsjdp3E-_gYF2pbUibDjPWAC1a6c"

CLASS_NAMES = ["قرحة (Granulation)", "Slough", "نخر (Necrosis)"]
CLASS_COLORS = {
    "قرحة (Granulation)": (255, 0, 0),   # أحمر
    "Slough": (255, 255, 0),             # أصفر
    "نخر (Necrosis)": (0, 0, 0)          # أسود
}

segmenter = None

# =========================================================
# تحميل النموذج
# =========================================================
def initialize_model():
    """تحميل نموذج Unet EfficientNet من Google Drive"""
    global segmenter

    if not os.path.exists(MODEL_PATH):
        print("📥 تحميل النموذج من Google Drive...")
        gdown.download(MODEL_URL, MODEL_PATH, quiet=False)

    try:
        print("🔄 تحميل Unet EfficientNet...")
        model = smp.Unet(
            encoder_name="efficientnet-b0",
            encoder_weights=None,
            classes=len(CLASS_NAMES),
            activation="sigmoid"
        )

        checkpoint = torch.load(MODEL_PATH, map_location=DEVICE)
        if "state_dict" in checkpoint:
            state_dict = checkpoint["state_dict"]
        else:
            state_dict = checkpoint

        clean_state = {k.replace("module.", "").replace("model.", ""): v for k, v in state_dict.items()}
        model.load_state_dict(clean_state, strict=False)
        model.to(DEVICE)
        model.eval()
        segmenter = model
        print("✅ تم تحميل النموذج بنجاح.")
    except Exception as e:
        print(f"❌ فشل تحميل النموذج: {e}")
        import traceback; traceback.print_exc()
        segmenter = None

# =========================================================
# أدوات مساعدة
# =========================================================
def ensure_rgb(np_img):
    """تحويل الصورة إلى RGB إذا لزم"""
    if np_img.ndim == 2:
        return cv2.cvtColor(np_img, cv2.COLOR_GRAY2RGB)
    if np_img.shape[-1] == 4:
        return cv2.cvtColor(np_img, cv2.COLOR_RGBA2RGB)
    return np_img

def preprocess_image(img: Image.Image):
    img_np = ensure_rgb(np.array(img))
    img_resized = cv2.resize(img_np, (IMG_SIZE, IMG_SIZE))
    img_norm = img_resized.astype(np.float32) / 255.0

    # ✅ تطبيع ImageNet الصحيح
    mean = np.array([0.485, 0.456, 0.406])
    std  = np.array([0.229, 0.224, 0.225])
    img_norm = (img_norm - mean) / std

    # ✅ تحويل إلى double لأن النموذج يستخدم float64
    tensor = torch.from_numpy(img_norm).permute(2, 0, 1).unsqueeze(0).double()
    return tensor.to(DEVICE), img_np

# =========================================================
# التجزئة والتحليل
# =========================================================
def analyze_image(img: Image.Image):
    """تحليل صورة القدم وعرض النسب"""
    if segmenter is None:
        return img, img, {"خطأ": "النموذج غير مهيأ بعد."}

    try:
        print("🔍 بدء التحليل...")
        tensor, img_np = preprocess_image(img)

        with torch.no_grad():
            output = segmenter(tensor).cpu().squeeze(0).numpy()  # (3,H,W)

        masks = (output >= THRESHOLD).astype(np.uint8)

        # تنظيف الأقنعة
        kernel = np.ones((5,5), np.uint8)
        for i in range(masks.shape[0]):
            masks[i] = cv2.morphologyEx(masks[i], cv2.MORPH_OPEN, kernel)
            masks[i] = cv2.morphologyEx(masks[i], cv2.MORPH_CLOSE, kernel)

        # حساب النسب
        total_pixels = masks.shape[1] * masks.shape[2]
        ratios = {
            CLASS_NAMES[0]: np.sum(masks[0]) / total_pixels * 100,
            CLASS_NAMES[1]: np.sum(masks[1]) / total_pixels * 100,
            CLASS_NAMES[2]: np.sum(masks[2]) / total_pixels * 100
        }
        total_ratio = sum(ratios.values())

        # إنشاء قناع لوني
        color_mask = np.zeros((masks.shape[1], masks.shape[2], 3), dtype=np.uint8)
        color_mask[masks[0] == 1] = CLASS_COLORS[CLASS_NAMES[0]]
        color_mask[masks[1] == 1] = CLASS_COLORS[CLASS_NAMES[1]]
        color_mask[masks[2] == 1] = CLASS_COLORS[CLASS_NAMES[2]]

        color_mask = cv2.resize(color_mask, (img_np.shape[1], img_np.shape[0]))

        # دمج القناع مع الصورة
        alpha = 0.5
        blended = cv2.addWeighted(img_np, 1 - alpha, color_mask, alpha, 0)

        # تقييم الخطورة
        if total_ratio == 0:
            risk = "No Risk 🟢"
        elif total_ratio < 1:
            risk = "Low Risk 🟡"
        elif total_ratio < 5:
            risk = "Medium Risk 🟠"
        else:
            risk = "High Risk 🔴"

        report = {
            "نسب الأنسجة (%)": {k: f"{v:.2f}" for k, v in ratios.items()},
            "إجمالي (%)": f"{total_ratio:.2f}",
            "مستوى الخطورة": risk
        }

        print(f"📊 النتائج: {report}")
        return Image.fromarray(blended), Image.fromarray(color_mask), report

    except Exception as e:
        print(f"❌ خطأ أثناء التحليل: {e}")
        import traceback; traceback.print_exc()
        return img, img, {"خطأ": str(e)}

# =========================================================
# واجهة Gradio
# =========================================================
def build_ui():
    with gr.Blocks(title="تحليل قرحة القدم - EfficientNet Unet", theme=gr.themes.Soft()) as demo:
        gr.Markdown("# 🦶 تحليل صورة القدم السكري (Unet + EfficientNet)")
        gr.Markdown("الكشف عن أنواع الأنسجة المصابة (قرحة / Slough / نخر) وتقدير مستوى الخطورة.")

        with gr.Row():
            with gr.Column(scale=1):
                input_img = gr.Image(type="pil", label="📤 ارفع صورة القدم", height=320)
                analyze_btn = gr.Button("🔍 بدء التحليل", variant="primary")

            with gr.Column(scale=1):
                out_blended = gr.Image(type="pil", label="🩸 الصورة مع القناع", height=320)
                out_mask = gr.Image(type="pil", label="🧩 القناع اللوني", height=320)
                out_json = gr.JSON(label="📊 التقرير التفصيلي")

        analyze_btn.click(
            fn=analyze_image,
            inputs=[input_img],
            outputs=[out_blended, out_mask, out_json]
        )
    return demo

# =========================================================
# تشغيل التطبيق
# =========================================================
if __name__ == "__main__":
    print("🚀 تهيئة النموذج...")
    initialize_model()
    app = build_ui()
    app.launch(server_name="0.0.0.0", server_port=7860)