DiabeticRetinpathyClassifier

Diabetic Retinopathy Grading Model (V2)

This is a multi-task deep learning model trained to classify the severity of Diabetic Retinopathy (DR) from retinal fundus images. It is based on the EfficientNet-B3 architecture and was specifically optimized to improve the Quadratic Weighted Kappa (QWK) score, a clinically relevant metric for ordinal classification tasks like DR grading.

This model is the second iteration (V2) of a project focused on building a diagnostically "smarter" classifier that is more sensitive to severe, vision-threatening stages of the disease.

Model Details

• Architecture: timm/efficientnet_b3 backbone with a custom multi-task head.
• Input Size: 512x512 pixels.
• Output: A dictionary containing logits for three tasks:
  • severity: 5 classes (0: No DR, 1: Mild, 2: Moderate, 3: Severe, 4: Proliferative).
  • lesions: 5 classes (multi-label for various lesion types).
  • regions: 5 classes (multi-label for affected anatomical regions).
• Training Objective: The model was trained focusing only on the severity task by setting the loss weights for auxiliary tasks to zero. The auxiliary heads can still produce outputs for interpretability.

How to Get Started & Use

The model can be easily loaded from Hugging Face Hub for inference.

# Install required libraries
pip install torch torchvision timm albumentations huggingface-hub numpy pillow opencv-python

import torch
import torch.nn as nn
import torch.nn.functional as F
import timm
from PIL import Image
import numpy as np
import albumentations as A
from albumentations.pytorch import ToTensorV2
from huggingface_hub import hf_hub_download

# Define the model architecture
class MultiTaskDRModel(nn.Module):
    def __init__(self, model_name='efficientnet_b3', num_classes=5,
                 num_lesion_types=5, num_regions=5, pretrained=False):
        super(MultiTaskDRModel, self).__init__()
        self.backbone = timm.create_model(model_name, pretrained=pretrained, num_classes=0)
        self.feature_dim = self.backbone.num_features
        
        self.attention = nn.Sequential(
            nn.AdaptiveAvgPool2d(1), nn.Flatten(),
            nn.Linear(self.feature_dim, self.feature_dim // 8), nn.ReLU(inplace=True),
            nn.Linear(self.feature_dim // 8, self.feature_dim), nn.Sigmoid()
        )
        
        self.feature_norm = nn.BatchNorm1d(self.feature_dim)
        self.dropout = nn.Dropout(0.4)
        
        self.severity_classifier = nn.Sequential(
            nn.Linear(self.feature_dim, self.feature_dim // 2), nn.ReLU(inplace=True),
            nn.Dropout(0.2), nn.Linear(self.feature_dim // 2, num_classes)
        )
        
        self.lesion_detector = nn.Sequential(
            nn.Linear(self.feature_dim, self.feature_dim // 4), nn.ReLU(inplace=True),
            nn.Dropout(0.2), nn.Linear(self.feature_dim // 4, num_lesion_types)
        )
        
        self.region_predictor = nn.Sequential(
            nn.Linear(self.feature_dim, self.feature_dim // 4), nn.ReLU(inplace=True),
            nn.Dropout(0.2), nn.Linear(self.feature_dim // 4, num_regions)
        )
    
    def forward(self, x):
        features = self.backbone.forward_features(x)
        pooled_features = F.adaptive_avg_pool2d(features, 1).flatten(1)
        attention_weights = self.attention(pooled_features.unsqueeze(-1).unsqueeze(-1))
        features = pooled_features * attention_weights
        features = self.feature_norm(features)
        features = self.dropout(features)
        
        severity_logits = self.severity_classifier(features)
        lesion_logits = self.lesion_detector(features)
        region_logits = self.region_predictor(features)
        
        return {
            'severity': severity_logits,
            'lesions': lesion_logits,
            'regions': region_logits,
            'features': features
        }

# Load the model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = MultiTaskDRModel()

# Download and load the checkpoint
model_path = hf_hub_download(
    repo_id="dheeren-tejani/DiabeticRetinpathyClassifier",
    filename="best_model_v2.pth"
)
checkpoint = torch.load(model_path, map_location=device, weights_only=False)
model.load_state_dict(checkpoint['model_state_dict'])
model.to(device)
model.eval()

print("Model loaded successfully!")

# Preprocessing function
def preprocess_image(image_path):
    transforms = A.Compose([
        A.Resize(512, 512),
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ToTensorV2(),
    ])
    image = np.array(Image.open(image_path).convert("RGB"))
    image_tensor = transforms(image=image)['image'].unsqueeze(0)
    return image_tensor

# Example inference
def predict_dr_severity(image_path):
    image_tensor = preprocess_image(image_path).to(device)
    
    with torch.no_grad():
        outputs = model(image_tensor)
    
    # Get severity prediction
    severity_probs = torch.softmax(outputs['severity'], dim=1)
    predicted_class = torch.argmax(severity_probs, dim=1).item()
    confidence = severity_probs[0, predicted_class].item()
    
    severity_labels = {
        0: "No DR",
        1: "Mild DR", 
        2: "Moderate DR",
        3: "Severe DR",
        4: "Proliferative DR"
    }
    
    return {
        'predicted_severity': severity_labels[predicted_class],
        'confidence': confidence,
        'all_probabilities': severity_probs[0].cpu().numpy()
    }

# Example usage
# result = predict_dr_severity("path/to/your/fundus_image.jpg")
# print(f"Predicted: {result['predicted_severity']} (Confidence: {result['confidence']:.3f})")

Training Details

V2 Improvements

This model (V2) was specifically designed to address the shortcomings of a baseline model (V1) that struggled with severe-stage DR detection:

• Higher Resolution: Increased from 224×224 to 512×512 to capture finer pathological details
• Class Balancing: Implemented WeightedRandomSampler to oversample rare minority classes (Severe and Proliferative DR)
• Focal Loss: Replaced standard Cross-Entropy with Focal Loss (γ=2.0) to focus on hard-to-classify examples
• Focused Training: Set auxiliary task weights to zero, dedicating full model capacity to severity classification

Hyperparameters

• Optimizer: AdamW
• Learning Rate: 1e-4
• Scheduler: CosineAnnealingWarmRestarts (T_MAX=10)
• Batch Size: 16
• Epochs: 17 (Early stopping)
• Image Size: 512×512

Performance

The model was evaluated on a held-out validation set of 735 images:

Key Achievement

The V2 model achieved a +3.5% improvement in QWK over the V1 baseline (0.761), indicating it makes "smarter" errors that are more aligned with clinical judgment, despite lower overall accuracy. This trade-off prioritizes clinically relevant performance over naive accuracy.

Limitations

⚠️ Important Disclaimers:

• This model was trained on a single public dataset and may not generalize to different clinical settings, camera types, or patient demographics
• The dataset may contain inherent demographic biases
• This is NOT a medical device and should not be used for actual clinical diagnosis
• Always consult qualified healthcare professionals for medical decisions

Citation

If you use this model in your research, please cite:

@misc{dheerentejani2025dr,
  author = {Dheeren Tejani},
  title = {Diabetic Retinopathy Grading Model V2},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
  howpublished = {\url{https://huggingface.co/dheeren-tejani/DiabeticRetinpathyClassifier}},
}

License

This model is released under the Apache 2.0 License.