import gradio as gr import torch import torch.nn as nn import timm import cv2 import numpy as np import joblib import json import shap import tensorflow as tf import matplotlib.pyplot as plt import io from PIL import Image print("🚀 Initializing Final AI Interface (Hugging Face Mode)...") device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') classes = ["Normal", "Cyst", "Tumor"] # ========================================== # 1. إعداد المسارات (المسارات المحلية في Hugging Face) # ========================================== VIT_MODEL_PATH = 'best_vit_model.pth' FUSION_MODEL_PATH = 'best_vit_fusion_model.h5' SCALER_PATH = 'clinical_scaler.pkl' SHAP_BG_PATH = 'shap_background_100.npy' JSON_CASES_PATH = 'clinical_data.json' # ========================================== # 2. تحميل النماذج والملفات المرجعية # ========================================== print("📦 Loading Models and Assets...") # أ- تحميل نموذج ViT vit_model = timm.create_model('vit_base_patch16_224', pretrained=False, num_classes=3) vit_model.load_state_dict(torch.load(VIT_MODEL_PATH, map_location=device)) vit_model.head = nn.Identity() vit_model.to(device) vit_model.eval() # ب- تحميل نموذج الدمج (Fusion) fusion_model = tf.keras.models.load_model(FUSION_MODEL_PATH) # ج- تحميل مُقيس البيانات السريرية (Scaler) scaler = joblib.load(SCALER_PATH) # د- تحميل عينات SHAP try: background_data = np.load(SHAP_BG_PATH) explainer = shap.DeepExplainer(fusion_model, background_data) print("✅ SHAP Initialized with Saved Background.") except FileNotFoundError: print(f"❌ Error: Could not find {SHAP_BG_PATH}. Please generate and upload it first.") # ========================================== # 3. تحميل العينات الجاهزة للواجهة (Examples) # ========================================== try: with open(JSON_CASES_PATH, 'r') as f: saved_cases = json.load(f) examples_img_only = [[c["image_path"]] for c in saved_cases] examples_fusion = [[ c["image_path"], c["age"], c["bmi"], c["rad_size"], c["gender"], c["smoking"], c["alcohol"], c["egfr"], c["ckd"], c["diabetes"] ] for c in saved_cases] print(f"✅ Loaded {len(saved_cases)} examples for UI.") except FileNotFoundError: print(f"⚠️ Warning: Could not find JSON file at {JSON_CASES_PATH}. Examples won't be shown.") examples_img_only = [] examples_fusion = [] # ========================================== # 4. دوال التوقع والتفسير # ========================================== def fig_to_img(fig): buf = io.BytesIO() fig.savefig(buf, format='png', bbox_inches='tight', dpi=100) buf.seek(0) img = Image.open(buf) plt.close(fig) return img def predict_image_only(image): if image is None: return None img_resized = cv2.resize(image, (224, 224)) img_normalized = img_resized.astype(np.float32) / 255.0 mean = np.array([0.485, 0.456, 0.406]); std = np.array([0.229, 0.224, 0.225]) img_tensor = (img_normalized - mean) / std img_tensor = torch.tensor(np.transpose(img_tensor, (2, 0, 1)), dtype=torch.float32).unsqueeze(0).to(device) with torch.no_grad(): img_features = vit_model(img_tensor).cpu().numpy() dummy_clinical = np.zeros((1, 10)) X_fusion = np.concatenate([img_features, dummy_clinical], axis=1) preds = fusion_model.predict(X_fusion, verbose=0)[0] return {classes[i]: float(preds[i]) for i in range(3)} def predict_fusion(image, age, bmi, rad_size, gender, smoking, alcohol, egfr, ckd, diabetes): if image is None: return None, None, None img_resized = cv2.resize(image, (224, 224)) img_normalized = img_resized.astype(np.float32) / 255.0 mean = np.array([0.485, 0.456, 0.406]); std = np.array([0.229, 0.224, 0.225]) img_tensor = (img_normalized - mean) / std img_tensor = torch.tensor(np.transpose(img_tensor, (2, 0, 1)), dtype=torch.float32).unsqueeze(0).to(device) with torch.no_grad(): img_features = vit_model(img_tensor).cpu().numpy() g_map = {"Male": 0, "Female": 1, "Other": 2} s_map = {"Never": 0, "Previous Smoker": 1, "Current Smoker": 2} a_map = {"None": 0, "Moderate": 1, "Heavy": 2} comorb_score = int(ckd) + int(diabetes) clin_data = np.array([[age, bmi, rad_size, g_map[gender], s_map[smoking], a_map[alcohol], egfr, int(ckd), int(diabetes), comorb_score]]) clin_scaled = scaler.transform(clin_data) X_fusion = np.concatenate([img_features, clin_scaled], axis=1) preds = fusion_model.predict(X_fusion, verbose=0)[0] pred_class = np.argmax(preds) result_dict = {classes[i]: float(preds[i]) for i in range(3)} shap_vals = explainer.shap_values(X_fusion) if isinstance(shap_vals, list): shap_patient = shap_vals[pred_class][0] elif len(shap_vals.shape) == 3: shap_patient = shap_vals[0, :, pred_class] else: shap_patient = shap_vals[0] img_shap_sum = np.sum(np.abs(shap_patient[:768])) clin_shap_sum = np.sum(np.abs(shap_patient[768:])) total_shap = img_shap_sum + clin_shap_sum img_pct = (img_shap_sum / total_shap) * 100 if total_shap > 0 else 0 clin_pct = (clin_shap_sum / total_shap) * 100 if total_shap > 0 else 0 fig_pie, ax_pie = plt.subplots(figsize=(4, 4)) ax_pie.pie([img_pct, clin_pct], labels=['Image (ViT)', 'Clinical'], autopct='%1.1f%%', colors=['#ff9999', '#66b3ff'], explode=(0.1, 0), shadow=True, startangle=140) ax_pie.set_title("Modality Contribution") img_pie = fig_to_img(fig_pie) clin_names = np.array(['Age', 'BMI', 'Rad Size', 'Gender', 'Smoking', 'Alcohol', 'eGFR', 'CKD', 'Diabetes', 'Comorb Score']) clin_impact = np.abs(shap_patient[768:]) sorted_idx = np.argsort(clin_impact)[-5:] fig_bar, ax_bar = plt.subplots(figsize=(5, 4)) ax_bar.barh(clin_names[sorted_idx], clin_impact[sorted_idx], color='#66b3ff', edgecolor='black') ax_bar.set_title(f"Top Clinical Factors for {classes[pred_class]}") ax_bar.set_xlabel("Absolute SHAP Impact") plt.tight_layout() img_bar = fig_to_img(fig_bar) return result_dict, img_pie, img_bar # ========================================== # 5. واجهة المستخدم (UI) # ========================================== print("🌐 Launching Gradio Interface...") with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.Markdown("
🩺 إعداد حعفر شهيرة - سالي حليبية - سلاف خدام
") with gr.Tabs(): with gr.TabItem("🖼️ Scenario 1: Image Only"): with gr.Row(): with gr.Column(): img_only_in = gr.Image(label="Upload CT Scan", type="numpy") btn_only = gr.Button("Analyze Image", variant="primary") if examples_img_only: gr.Examples(examples=examples_img_only, inputs=img_only_in) with gr.Column(): out_only_label = gr.Label(label="Prediction (Ignoring Clinical Data)") btn_only.click(predict_image_only, inputs=img_only_in, outputs=out_only_label) with gr.TabItem("🧬 Scenario 2: Image + Clinical Fusion"): with gr.Row(): with gr.Column(): img_fusion_in = gr.Image(label="CT Scan", type="numpy") with gr.Row(): age_in = gr.Number(label="Age", value=50) bmi_in = gr.Number(label="BMI", value=25.0) with gr.Row(): rad_size_in = gr.Number(label="Radiographic Size (cm)", value=3.5) egfr_in = gr.Number(label="eGFR (Kidney Function)", value=90.0) gender_in = gr.Radio(["Male", "Female", "Other"], label="Gender", value="Male") smoking_in = gr.Radio(["Never", "Previous Smoker", "Current Smoker"], label="Smoking", value="Never") alcohol_in = gr.Radio(["None", "Moderate", "Heavy"], label="Alcohol Use", value="None") with gr.Row(): ckd_in = gr.Checkbox(label="Chronic Kidney Disease (CKD)") diabetes_in = gr.Checkbox(label="Diabetes") btn_fusion = gr.Button("Analyze Fusion Data", variant="primary") if examples_fusion: gr.Examples(examples=examples_fusion, inputs=[img_fusion_in, age_in, bmi_in, rad_size_in, gender_in, smoking_in, alcohol_in, egfr_in, ckd_in, diabetes_in]) with gr.Column(): out_fusion_label = gr.Label(label="Final AI Diagnosis") gr.Markdown("### 🔍 AI Explainability (Why did the model decide this?)") with gr.Row(): out_pie = gr.Image(label="Modality Impact") out_bar = gr.Image(label="Top Clinical Factors") btn_fusion.click( fn=predict_fusion, inputs=[img_fusion_in, age_in, bmi_in, rad_size_in, gender_in, smoking_in, alcohol_in, egfr_in, ckd_in, diabetes_in], outputs=[out_fusion_label, out_pie, out_bar] ) # تشغيل الواجهة if __name__ == "__main__": demo.launch()