Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,189 +1,486 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
# Ready for HuggingFace Spaces
|
| 4 |
-
# =============================================
|
| 5 |
-
import os
|
| 6 |
-
import io
|
| 7 |
-
import cv2
|
| 8 |
-
import json
|
| 9 |
import torch
|
|
|
|
| 10 |
import numpy as np
|
| 11 |
import gradio as gr
|
| 12 |
-
import torch.nn as nn
|
| 13 |
-
import torchvision.transforms as transforms
|
| 14 |
-
from torchvision import models
|
| 15 |
from PIL import Image
|
| 16 |
-
import
|
|
|
|
|
|
|
|
|
|
| 17 |
import albumentations as A
|
| 18 |
-
|
| 19 |
-
import
|
| 20 |
-
from
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
clf_model_path = '/content/drive/MyDrive/models/efficientnetb3_dfu_model.keras'
|
| 43 |
-
clf_model = load_model(clf_model_path)
|
| 44 |
-
|
| 45 |
-
return seg_model, clf_model, device
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
# =============================================
|
| 49 |
-
# Model Definitions (for optional retraining)
|
| 50 |
-
# =============================================
|
| 51 |
-
|
| 52 |
-
class FuSegNet(nn.Module):
|
| 53 |
-
def __init__(self):
|
| 54 |
-
super().__init__()
|
| 55 |
-
self.model = smp.Unet('efficientnet-b7', encoder_weights='imagenet', in_channels=3, classes=1)
|
| 56 |
-
|
| 57 |
def forward(self, x):
|
| 58 |
-
return self.
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
upper_red1 = np.array([10, 255, 255])
|
| 72 |
-
lower_red2 = np.array([160,
|
| 73 |
-
upper_red2 = np.array([
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
else:
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
images = [img1, img2]
|
| 109 |
-
|
|
|
|
| 110 |
|
| 111 |
-
for img in images:
|
|
|
|
|
|
|
| 112 |
if img is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 113 |
continue
|
| 114 |
-
temp_path = '/tmp/temp_img.jpg'
|
| 115 |
-
img.save(temp_path)
|
| 116 |
-
|
| 117 |
-
# Classification (TensorFlow)
|
| 118 |
-
img_tf = img.resize((300, 300))
|
| 119 |
-
img_arr = np.array(img_tf) / 255.0
|
| 120 |
-
pred = clf_model.predict(np.expand_dims(img_arr, axis=0))[0][0]
|
| 121 |
-
classification = 'Abnormal (Ulcer)' if pred > 0.5 else 'Normal (Healthy)'
|
| 122 |
-
|
| 123 |
-
# Segmentation (PyTorch)
|
| 124 |
-
transform = A.Compose([
|
| 125 |
-
A.Resize(512, 512),
|
| 126 |
-
A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
|
| 127 |
-
ToTensorV2()
|
| 128 |
-
])
|
| 129 |
-
|
| 130 |
-
image_np = np.array(img.convert('RGB'))
|
| 131 |
-
image_tensor = transform(image=image_np)['image'].unsqueeze(0).to(device)
|
| 132 |
-
|
| 133 |
-
with torch.no_grad():
|
| 134 |
-
mask_pred = torch.sigmoid(seg_model(image_tensor)).cpu().numpy()[0, 0]
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
)
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
#
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import tensorflow as tf
|
| 2 |
+
from tensorflow.keras.models import load_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
import torch
|
| 4 |
+
import torch.nn as nn
|
| 5 |
import numpy as np
|
| 6 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
| 7 |
from PIL import Image
|
| 8 |
+
import cv2
|
| 9 |
+
from torchvision import transforms
|
| 10 |
+
import gdown
|
| 11 |
+
import os
|
| 12 |
import albumentations as A
|
| 13 |
+
import segmentation_models_pytorch as smp
|
| 14 |
+
import requests
|
| 15 |
+
from io import BytesIO
|
| 16 |
+
|
| 17 |
+
# تعريفات عالمية
|
| 18 |
+
classifier = None
|
| 19 |
+
segmenter = None
|
| 20 |
+
class_names = ["Abnormal(Ulcer)", "Normal(Healthy skin)"]
|
| 21 |
+
IMG_SIZE = 224
|
| 22 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 23 |
+
|
| 24 |
+
class FUSegNet(nn.Module):
|
| 25 |
+
"""FUSegNet نموذج مخصص لمطابقة هيكل"""
|
| 26 |
+
def __init__(self, encoder_name='efficientnet-b7', classes=1, activation='sigmoid'):
|
| 27 |
+
super(FUSegNet, self).__init__()
|
| 28 |
+
|
| 29 |
+
self.unet = smp.Unet(
|
| 30 |
+
encoder_name=encoder_name,
|
| 31 |
+
encoder_weights=None,
|
| 32 |
+
classes=classes,
|
| 33 |
+
activation=activation,
|
| 34 |
+
decoder_attention_type='pscse', # إضافة نوع الانتباه المخصص
|
| 35 |
+
)
|
| 36 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
def forward(self, x):
|
| 38 |
+
return self.unet(x)
|
| 39 |
+
|
| 40 |
+
def check_and_download(url, path, min_size_mb=50):
|
| 41 |
+
"""التحقق من وجود الملفات وتحميلها إذا كانت مفقودة"""
|
| 42 |
+
if not os.path.exists(path) or os.path.getsize(path) < min_size_mb * 1024 * 1024:
|
| 43 |
+
print(f"📥 تحميل النموذج من Google Drive: {url}")
|
| 44 |
+
try:
|
| 45 |
+
gdown.download(url, path, quiet=False, fuzzy=True)
|
| 46 |
+
if os.path.exists(path):
|
| 47 |
+
size = os.path.getsize(path) / (1024 * 1024)
|
| 48 |
+
print(f"✅ تم التحميل بنجاح: {os.path.basename(path)} ({size:.2f} MB)")
|
| 49 |
+
else:
|
| 50 |
+
print(f"❌ فشل التحميل: {path}")
|
| 51 |
+
except Exception as e:
|
| 52 |
+
print(f"❌ خطأ في التحميل: {e}")
|
| 53 |
+
else:
|
| 54 |
+
size = os.path.getsize(path) / (1024 * 1024)
|
| 55 |
+
print(f"✅ الملف موجود: {os.path.basename(path)} ({size:.2f} MB)")
|
| 56 |
+
|
| 57 |
+
def initialize_models():
|
| 58 |
+
"""تهيئة النماذج"""
|
| 59 |
+
global classifier, segmenter
|
| 60 |
+
|
| 61 |
+
# روابط ومسارات النماذج
|
| 62 |
+
EFF_MODEL_URL = "https://drive.google.com/uc?id=1vVmA_-D3pZPbKHrPFEbDJd2nexF1H9Ni"
|
| 63 |
+
SEG_MODEL_URL = "https://drive.google.com/uc?id=13jMlcH9yTSejL_IfMDXqdAiVy6Z_SpE1"
|
| 64 |
+
|
| 65 |
+
EFF_MODEL_PATH = "efficientnetb3_dfu_model.keras"
|
| 66 |
+
SEG_MODEL_PATH = "best_model.pth"
|
| 67 |
+
|
| 68 |
+
# تحميل النماذج
|
| 69 |
+
check_and_download(EFF_MODEL_URL, EFF_MODEL_PATH, min_size_mb=50)
|
| 70 |
+
check_and_download(SEG_MODEL_URL, SEG_MODEL_PATH, min_size_mb=100)
|
| 71 |
+
|
| 72 |
+
# ------ نموذج التصنيف (Keras)
|
| 73 |
+
try:
|
| 74 |
+
print("🔄 ��اري تحميل نموذج التصنيف...")
|
| 75 |
+
classifier = tf.keras.models.load_model(EFF_MODEL_PATH, compile=False)
|
| 76 |
+
print("✅ تم تحميل نموذج التصنيف بنجاح!")
|
| 77 |
+
except Exception as e:
|
| 78 |
+
print(f"❌ خطأ أثناء تحميل نموذج التصنيف: {e}")
|
| 79 |
+
classifier = None
|
| 80 |
+
|
| 81 |
+
# ------ نموذج التجزئة (FUSegNet)
|
| 82 |
+
try:
|
| 83 |
+
print("🔄 جاري تحميل نموذج FUSegNet...")
|
| 84 |
+
|
| 85 |
+
# بناء النموذج بنفس مواصفات التدريب
|
| 86 |
+
segmenter = FUSegNet(
|
| 87 |
+
encoder_name='efficientnet-b7',
|
| 88 |
+
classes=1,
|
| 89 |
+
activation='sigmoid'
|
| 90 |
+
)
|
| 91 |
+
|
| 92 |
+
# تحميل الأوزان
|
| 93 |
+
checkpoint = torch.load(SEG_MODEL_PATH, map_location=DEVICE)
|
| 94 |
+
|
| 95 |
+
# معالجة مرنة للأوزان
|
| 96 |
+
if 'state_dict' in checkpoint:
|
| 97 |
+
state_dict = checkpoint['state_dict']
|
| 98 |
+
else:
|
| 99 |
+
state_dict = checkpoint
|
| 100 |
+
|
| 101 |
+
# تنظيف مفاتيح state_dict
|
| 102 |
+
new_state_dict = {}
|
| 103 |
+
for k, v in state_dict.items():
|
| 104 |
+
# إزالة البادئات المختلفة
|
| 105 |
+
new_key = k.replace('module.', '').replace('model.', '').replace('unet.', '')
|
| 106 |
+
new_state_dict[new_key] = v
|
| 107 |
+
|
| 108 |
+
# تحميل الأوزان مع التعامل مع المفاتيح المفقودة
|
| 109 |
+
model_dict = segmenter.state_dict()
|
| 110 |
+
|
| 111 |
+
# 1. محاولة التحميل المباشر أولاً
|
| 112 |
+
try:
|
| 113 |
+
segmenter.load_state_dict(new_state_dict, strict=True)
|
| 114 |
+
print("✅ تم تحميل الأوزان بنجاح (وضع صارم)")
|
| 115 |
+
except:
|
| 116 |
+
# 2. إذا فشل، حاول التحميل المرن
|
| 117 |
+
matched_keys = []
|
| 118 |
+
missing_keys = []
|
| 119 |
+
unexpected_keys = []
|
| 120 |
+
|
| 121 |
+
for name, param in model_dict.items():
|
| 122 |
+
if name in new_state_dict:
|
| 123 |
+
if param.shape == new_state_dict[name].shape:
|
| 124 |
+
param.data.copy_(new_state_dict[name])
|
| 125 |
+
matched_keys.append(name)
|
| 126 |
+
else:
|
| 127 |
+
missing_keys.append(name)
|
| 128 |
+
else:
|
| 129 |
+
missing_keys.append(name)
|
| 130 |
+
|
| 131 |
+
for key in new_state_dict:
|
| 132 |
+
if key not in model_dict:
|
| 133 |
+
unexpected_keys.append(key)
|
| 134 |
+
|
| 135 |
+
print(f"📊 إحصائيات التحميل:")
|
| 136 |
+
print(f" - المفاتيح المتطابقة: {len(matched_keys)}")
|
| 137 |
+
print(f" - المفاتيح المفقودة: {len(missing_keys)}")
|
| 138 |
+
print(f" - المفاتيح غير المتوقعة: {len(unexpected_keys)}")
|
| 139 |
+
|
| 140 |
+
if len(matched_keys) > 0:
|
| 141 |
+
print("✅ تم تحميل بعض الأوزان بنجاح")
|
| 142 |
+
else:
|
| 143 |
+
print("❌ فشل تحميل الأوزان، استخدام النموذج بدون أوزان")
|
| 144 |
+
|
| 145 |
+
segmenter.to(DEVICE)
|
| 146 |
+
segmenter.eval()
|
| 147 |
+
|
| 148 |
+
# اختبار سريع للنموذج
|
| 149 |
+
with torch.no_grad():
|
| 150 |
+
test_input = torch.randn(1, 3, IMG_SIZE, IMG_SIZE).to(DEVICE)
|
| 151 |
+
test_output = segmenter(test_input)
|
| 152 |
+
print(f"🧪 اختبار النموذج: [{test_output.min().item():.6f}, {test_output.max().item():.6f}]")
|
| 153 |
+
|
| 154 |
+
except Exception as e:
|
| 155 |
+
print(f"❌ خطأ في تحميل نموذج التجزئة: {e}")
|
| 156 |
+
segmenter = None
|
| 157 |
+
|
| 158 |
+
def get_preprocessing_fn():
|
| 159 |
+
"""دالة المعالجة المسبقة المتوافقة مع EfficientNet-B7"""
|
| 160 |
+
from segmentation_models_pytorch.encoders import get_preprocessing_fn
|
| 161 |
+
return get_preprocessing_fn('efficientnet-b7', pretrained='imagenet')
|
| 162 |
+
|
| 163 |
+
def smart_ulcer_detection(img: Image.Image):
|
| 164 |
+
"""كشف القرحة باستخدام معالجة الصور المتقدمة"""
|
| 165 |
+
print("🔍 استخدام الخوارزمية الذكية للكشف عن القرحة...")
|
| 166 |
+
|
| 167 |
+
img_np = np.array(img)
|
| 168 |
+
height, width = img_np.shape[:2]
|
| 169 |
+
|
| 170 |
+
# 1. تحويل إلى مساحات لون مختلفة
|
| 171 |
+
hsv = cv2.cvtColor(img_np, cv2.COLOR_RGB2HSV)
|
| 172 |
+
lab = cv2.cvtColor(img_np, cv2.COLOR_RGB2LAB)
|
| 173 |
+
|
| 174 |
+
# 2. كشف الألوان المرتبطة بالقرحة
|
| 175 |
+
# الأحمر النشط
|
| 176 |
+
lower_red1 = np.array([0, 60, 60])
|
| 177 |
upper_red1 = np.array([10, 255, 255])
|
| 178 |
+
lower_red2 = np.array([160, 60, 60])
|
| 179 |
+
upper_red2 = np.array([180, 255, 255])
|
| 180 |
+
red_mask = cv2.inRange(hsv, lower_red1, upper_red1) + cv2.inRange(hsv, lower_red2, upper_red2)
|
| 181 |
+
|
| 182 |
+
# البني/الأسود (أنسجة ميتة)
|
| 183 |
+
lower_brown = np.array([0, 40, 20])
|
| 184 |
+
upper_brown = np.array([20, 200, 150])
|
| 185 |
+
brown_mask = cv2.inRange(hsv, lower_brown, upper_brown)
|
| 186 |
+
|
| 187 |
+
# 3. كشف التغيرات في الإضاءة
|
| 188 |
+
l_channel = lab[:,:,0]
|
| 189 |
+
_, dark_areas = cv2.threshold(l_channel, 80, 255, cv2.THRESH_BINARY_INV)
|
| 190 |
+
|
| 191 |
+
# 4. دمج جميع الأقنعة
|
| 192 |
+
combined_mask = cv2.bitwise_or(red_mask, brown_mask)
|
| 193 |
+
combined_mask = cv2.bitwise_or(combined_mask, dark_areas)
|
| 194 |
+
|
| 195 |
+
# 5. تنظيف متقدم للقناع
|
| 196 |
+
kernel = np.ones((5,5), np.uint8)
|
| 197 |
+
combined_mask = cv2.morphologyEx(combined_mask, cv2.MORPH_OPEN, kernel)
|
| 198 |
+
combined_mask = cv2.morphologyEx(combined_mask, cv2.MORPH_CLOSE, kernel)
|
| 199 |
+
|
| 200 |
+
# 6. الاحتفاظ فقط بالمناطق الكبيرة
|
| 201 |
+
contours, _ = cv2.findContours(combined_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
| 202 |
+
final_mask = np.zeros_like(combined_mask)
|
| 203 |
+
|
| 204 |
+
min_area = height * width * 0.002 # 0.2% من مساحة الصورة
|
| 205 |
+
|
| 206 |
+
for contour in contours:
|
| 207 |
+
area = cv2.contourArea(contour)
|
| 208 |
+
if area > min_area:
|
| 209 |
+
perimeter = cv2.arcLength(contour, True)
|
| 210 |
+
if perimeter > 0:
|
| 211 |
+
circularity = 4 * np.pi * area / (perimeter * perimeter)
|
| 212 |
+
if circularity < 0.7: # استبعاد الأشكال الدائرية الكاملة
|
| 213 |
+
cv2.fillPoly(final_mask, [contour], 255)
|
| 214 |
+
|
| 215 |
+
ulcer_pixels = np.sum(final_mask > 0)
|
| 216 |
+
print(f"📊 الخوارزمية الذكية اكتشفت {ulcer_pixels} بيكسل قرحة")
|
| 217 |
+
|
| 218 |
+
return (final_mask > 0).astype(np.uint8)
|
| 219 |
+
|
| 220 |
+
def classify_image(img: Image.Image):
|
| 221 |
+
"""تصنيف الصورة"""
|
| 222 |
+
if classifier is None:
|
| 223 |
+
# استخدام كشف بسيط للون الأحمر للتصنيف الافتراضي
|
| 224 |
+
img_np = np.array(img)
|
| 225 |
+
hsv = cv2.cvtColor(img_np, cv2.COLOR_RGB2HSV)
|
| 226 |
+
red_mask = cv2.inRange(hsv, np.array([0,50,50]), np.array([10,255,255])) + \
|
| 227 |
+
cv2.inRange(hsv, np.array([160,50,50]), np.array([180,255,255]))
|
| 228 |
+
red_ratio = np.sum(red_mask > 0) / red_mask.size
|
| 229 |
+
result = "Abnormal(Ulcer)" if red_ratio > 0.005 else "Normal(Healthy skin)"
|
| 230 |
+
print(f"🎯 التصنيف الافتراضي: {result} (نسبة الأحمر: {red_ratio:.4f})")
|
| 231 |
+
return result
|
| 232 |
+
|
| 233 |
+
try:
|
| 234 |
+
# معالجة الصورة للتصنيف
|
| 235 |
+
img_processed = img.resize((IMG_SIZE, IMG_SIZE))
|
| 236 |
+
img_array = tf.keras.preprocessing.image.img_to_array(img_processed)
|
| 237 |
+
img_array = np.expand_dims(img_array, axis=0)
|
| 238 |
+
img_array = tf.keras.applications.efficientnet.preprocess_input(img_array)
|
| 239 |
+
|
| 240 |
+
# التوقع
|
| 241 |
+
preds = classifier.predict(img_array, verbose=0)
|
| 242 |
+
pred_class = np.argmax(preds, axis=1)[0]
|
| 243 |
+
confidence = np.max(preds)
|
| 244 |
+
|
| 245 |
+
result = class_names[pred_class]
|
| 246 |
+
print(f"🎯 تصنيف النموذج: {result} (ثقة: {confidence:.3f})")
|
| 247 |
+
return result
|
| 248 |
+
|
| 249 |
+
except Exception as e:
|
| 250 |
+
print(f"❌ خطأ في التصنيف: {e}")
|
| 251 |
+
return "Abnormal(Ulcer)"
|
| 252 |
+
|
| 253 |
+
def segment_image(img: Image.Image):
|
| 254 |
+
"""تجزئة الصورة باستخدام FUSegNet"""
|
| 255 |
+
if segmenter is not None:
|
| 256 |
+
try:
|
| 257 |
+
print("🔄 جاري تجزئة الصورة باستخدام FUSegNet...")
|
| 258 |
+
|
| 259 |
+
# معالجة الصورة بنفس طريقة التدريب
|
| 260 |
+
preprocessing_fn = get_preprocessing_fn()
|
| 261 |
+
|
| 262 |
+
img_np = np.array(img)
|
| 263 |
+
img_resized = cv2.resize(img_np, (IMG_SIZE, IMG_SIZE))
|
| 264 |
+
|
| 265 |
+
# تطبيق المعالجة المسبقة
|
| 266 |
+
processed_img = preprocessing_fn(img_resized)
|
| 267 |
+
|
| 268 |
+
# تحويل إلى tensor
|
| 269 |
+
img_tensor = torch.from_numpy(processed_img).permute(2, 0, 1).unsqueeze(0).float().to(DEVICE)
|
| 270 |
+
|
| 271 |
+
# التوقع
|
| 272 |
+
with torch.no_grad():
|
| 273 |
+
output = segmenter(img_tensor)
|
| 274 |
+
pred = output.squeeze().cpu().numpy()
|
| 275 |
+
|
| 276 |
+
print(f"📊 إخراج النموذج: [{pred.min():.6f}, {pred.max():.6f}]")
|
| 277 |
+
|
| 278 |
+
# تجربة عتبات مختلفة
|
| 279 |
+
best_threshold = 0.3
|
| 280 |
+
best_pixels = 0
|
| 281 |
+
|
| 282 |
+
for threshold in [0.2, 0.3, 0.4, 0.5]:
|
| 283 |
+
mask_bin = (pred >= threshold).astype(np.uint8)
|
| 284 |
+
mask_resized = cv2.resize(mask_bin, (img.width, img.height))
|
| 285 |
+
ulcer_pixels = np.sum(mask_resized)
|
| 286 |
+
|
| 287 |
+
if ulcer_pixels > best_pixels:
|
| 288 |
+
best_pixels = ulcer_pixels
|
| 289 |
+
best_threshold = threshold
|
| 290 |
+
|
| 291 |
+
if best_pixels > 0:
|
| 292 |
+
mask_bin = (pred >= best_threshold).astype(np.uint8)
|
| 293 |
+
mask_resized = cv2.resize(mask_bin, (img.width, img.height))
|
| 294 |
+
print(f"✅ FUSegNet اكتشف {best_pixels} بيكسل (threshold: {best_threshold})")
|
| 295 |
+
return mask_resized
|
| 296 |
+
else:
|
| 297 |
+
print("⚠️ FUSegNet لم يعط نتائج، استخدام الخوارزمية الذكية")
|
| 298 |
+
|
| 299 |
+
except Exception as e:
|
| 300 |
+
print(f"❌ خطأ في نموذج التجزئة: {e}")
|
| 301 |
+
|
| 302 |
+
# استخدام الخوارزمية الذكية كبديل
|
| 303 |
+
return smart_ulcer_detection(img)
|
| 304 |
+
|
| 305 |
+
def calculate_risk(mask):
|
| 306 |
+
"""حساب مستوى الخطورة"""
|
| 307 |
+
ulcer_pixels = np.sum(mask == 1)
|
| 308 |
+
total_pixels = mask.size
|
| 309 |
+
percent = (ulcer_pixels / total_pixels) * 100
|
| 310 |
+
|
| 311 |
+
print(f"📏 مساحة القرحة: {ulcer_pixels}/{total_pixels} بيكسل ({percent:.4f}%)")
|
| 312 |
+
|
| 313 |
+
if ulcer_pixels == 0:
|
| 314 |
+
return 0.0, "No Risk", 0
|
| 315 |
+
elif percent <= 1:
|
| 316 |
+
return percent, "Low Risk", 1
|
| 317 |
+
elif percent <= 5:
|
| 318 |
+
return percent, "Medium Risk", 3
|
| 319 |
else:
|
| 320 |
+
return percent, "High Risk", 5
|
| 321 |
+
|
| 322 |
+
def apply_mask_to_image(img: Image.Image, mask):
|
| 323 |
+
"""تطبيق القناع على الصورة"""
|
| 324 |
+
img_np = np.array(img)
|
| 325 |
+
|
| 326 |
+
ulcer_pixels = np.sum(mask == 1)
|
| 327 |
+
if ulcer_pixels == 0:
|
| 328 |
+
print("✅ لا توجد قرحة مكتشفة")
|
| 329 |
+
return img
|
| 330 |
+
|
| 331 |
+
print(f"🎨 تطبيق القناع على {ulcer_pixels} بيكسل")
|
| 332 |
+
|
| 333 |
+
# إنشاء قناع ملون
|
| 334 |
+
colored_mask = np.zeros_like(img_np)
|
| 335 |
+
colored_mask[mask == 1] = [255, 0, 0] # أحمر
|
| 336 |
+
|
| 337 |
+
# دمج مع الشفافية
|
| 338 |
+
result = cv2.addWeighted(img_np, 0.6, colored_mask, 0.4, 0)
|
| 339 |
+
|
| 340 |
+
# إضافة حدود حمراء
|
| 341 |
+
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
| 342 |
+
for contour in contours:
|
| 343 |
+
area = cv2.contourArea(contour)
|
| 344 |
+
if area > 20:
|
| 345 |
+
cv2.drawContours(result, [contour], -1, (255, 0, 0), 3)
|
| 346 |
+
|
| 347 |
+
return Image.fromarray(result)
|
| 348 |
+
|
| 349 |
+
def analyze_images(img1, img2):
|
| 350 |
+
"""تحليل صورتين فقط لتحسين الأداء"""
|
| 351 |
images = [img1, img2]
|
| 352 |
+
gallery_output = []
|
| 353 |
+
json_output = {}
|
| 354 |
|
| 355 |
+
for idx, img in enumerate(images):
|
| 356 |
+
image_key = f"image_{idx+1}"
|
| 357 |
+
|
| 358 |
if img is None:
|
| 359 |
+
gallery_output.append(None)
|
| 360 |
+
json_output[image_key] = {
|
| 361 |
+
"status": "No image uploaded",
|
| 362 |
+
"classification": "None",
|
| 363 |
+
"ulcer_percentage": 0.0,
|
| 364 |
+
"risk_level": "None",
|
| 365 |
+
"risk_score": 0
|
| 366 |
+
}
|
| 367 |
continue
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 368 |
|
| 369 |
+
try:
|
| 370 |
+
print(f"\n{'='*40}")
|
| 371 |
+
print(f"🖼️ معالجة الصورة {idx+1}...")
|
| 372 |
+
print(f"📐 حجم الصورة: {img.size}")
|
| 373 |
+
|
| 374 |
+
# التصنيف
|
| 375 |
+
classification = classify_image(img)
|
| 376 |
+
print(f"🎯 التصنيف: {classification}")
|
| 377 |
+
|
| 378 |
+
# التجزئة
|
| 379 |
+
mask = segment_image(img)
|
| 380 |
+
masked_img = apply_mask_to_image(img, mask)
|
| 381 |
+
ulcer_percent, risk_level, risk_score = calculate_risk(mask)
|
| 382 |
+
|
| 383 |
+
gallery_output.append(masked_img)
|
| 384 |
+
json_output[image_key] = {
|
| 385 |
+
"status": "Analyzed",
|
| 386 |
+
"classification": classification,
|
| 387 |
+
"ulcer_percentage": round(ulcer_percent, 4),
|
| 388 |
+
"risk_level": risk_level,
|
| 389 |
+
"risk_score": risk_score,
|
| 390 |
+
"ulcer_pixels": int(np.sum(mask == 1)),
|
| 391 |
+
"total_pixels": mask.size
|
| 392 |
+
}
|
| 393 |
+
|
| 394 |
+
except Exception as e:
|
| 395 |
+
print(f"❌ خطأ في الصورة {idx+1}: {e}")
|
| 396 |
+
gallery_output.append(img)
|
| 397 |
+
json_output[image_key] = {
|
| 398 |
+
"status": f"Error: {str(e)}",
|
| 399 |
+
"classification": "Error",
|
| 400 |
+
"ulcer_percentage": 0.0,
|
| 401 |
+
"risk_level": "Error",
|
| 402 |
+
"risk_score": 0
|
| 403 |
+
}
|
| 404 |
+
|
| 405 |
+
print(f"\n✅ تم معالجة {len([x for x in gallery_output if x is not None])} صور")
|
| 406 |
+
return gallery_output, json_output
|
| 407 |
+
|
| 408 |
+
# تهيئة النماذج
|
| 409 |
+
print("🚀 جاري تهيئة النماذج...")
|
| 410 |
+
initialize_models()
|
| 411 |
+
|
| 412 |
+
# واجهة Gradio
|
| 413 |
+
with gr.Blocks(title="Diabetic Foot Ulcer Analysis", theme=gr.themes.Soft()) as demo:
|
| 414 |
+
gr.Markdown("""
|
| 415 |
+
# 🦶 Diabetic Foot Ulcer Detection & Risk Analysis
|
| 416 |
+
|
| 417 |
+
**تحليل قرحة القدم السكري باستخدام الذكاء الاصطناعي المتقدم**
|
| 418 |
+
|
| 419 |
+
⚡ **مميزات النظام:**
|
| 420 |
+
- تحليل صورتين فقط لتسريع الأداء
|
| 421 |
+
- كشف ذكي للقرحة باستخدام نموذج FUSegNet المدرب
|
| 422 |
+
- خوارزمية بديلة ذكية إذا فشلت النماذج
|
| 423 |
+
- تقييم دقيق لمستوى الخطورة
|
| 424 |
+
""")
|
| 425 |
+
|
| 426 |
+
with gr.Row():
|
| 427 |
+
with gr.Column(scale=1):
|
| 428 |
+
gr.Markdown("### 📤 رفع الصور")
|
| 429 |
+
with gr.Row():
|
| 430 |
+
img1 = gr.Image(type="pil", label="صورة القدم الأولى", height=250)
|
| 431 |
+
img2 = gr.Image(type="pil", label="صورة القدم الثانية", height=250)
|
| 432 |
+
|
| 433 |
+
analyze_btn = gr.Button(
|
| 434 |
+
"🔍 بدء التحليل",
|
| 435 |
+
variant="primary",
|
| 436 |
+
size="lg"
|
| 437 |
)
|
| 438 |
+
|
| 439 |
+
gr.Markdown("""
|
| 440 |
+
**💡 نصائح للاستخدام:**
|
| 441 |
+
- اختر صور واضحة للقدم
|
| 442 |
+
- تجنب الصور المضغوطة كثيراً
|
| 443 |
+
- الصور الملونة تعطي نتائج أفضل
|
| 444 |
+
- تأكد من إضاءة جيدة للصورة
|
| 445 |
+
""")
|
| 446 |
+
|
| 447 |
+
with gr.Column(scale=1):
|
| 448 |
+
gr.Markdown("### 📊 النتائج")
|
| 449 |
+
gallery = gr.Gallery(
|
| 450 |
+
label="الصور المحللة",
|
| 451 |
+
columns=2,
|
| 452 |
+
height="auto",
|
| 453 |
+
object_fit="contain"
|
| 454 |
+
)
|
| 455 |
+
|
| 456 |
+
json_output = gr.JSON(
|
| 457 |
+
label="النتائج التفصيلية",
|
| 458 |
+
show_label=True
|
| 459 |
+
)
|
| 460 |
+
|
| 461 |
+
# تعليمات إضافية
|
| 462 |
+
gr.Markdown("""
|
| 463 |
+
---
|
| 464 |
+
**📋 تفسير النتائج:**
|
| 465 |
+
- **No Risk**: لا توجد قرحة مكتشفة
|
| 466 |
+
- **Low Risk**: مساحة القرحة ≤ 1%
|
| 467 |
+
- **Medium Risk**: مساحة القرحة ≤ 5%
|
| 468 |
+
- **High Risk**: مساحة القرحة > 5%
|
| 469 |
+
|
| 470 |
+
**🔴 المناطق الحمراء**: تشير إلى ��واقع القرحة المكتشفة
|
| 471 |
+
""")
|
| 472 |
+
|
| 473 |
+
analyze_btn.click(
|
| 474 |
+
fn=analyze_images,
|
| 475 |
+
inputs=[img1, img2],
|
| 476 |
+
outputs=[gallery, json_output]
|
| 477 |
+
)
|
| 478 |
+
|
| 479 |
+
if __name__ == "__main__":
|
| 480 |
+
print("🌐 Starting Gradio server...")
|
| 481 |
+
print("✅ النظام جاهز لتحليل صورتين")
|
| 482 |
+
demo.launch(
|
| 483 |
+
server_name="0.0.0.0",
|
| 484 |
+
server_port=7860,
|
| 485 |
+
share=True
|
| 486 |
+
)
|