Update README.md

metadata

license: apache-2.0
language:
  - en
metrics:
  - accuracy
base_model:
  - microsoft/resnet-50
  - timm/vgg19.tv_in1k
  - google/vit-base-patch16-224
  - xai-org/grok-1
pipeline_tag: image-classification
tags:
  - Ocular-Toxoplasmosis(FundusImages)
  - Retinal-images(Diabetics,Cataract,Gulocoma,Healthy)
  - Pytorch
  - Transformers
  - Image-Classification
  - Image_feature_extraction
  - Grad-CAM
  - XAI-Visualization

Model Card: ROYXAI [Vision Transformer + VGG19 + ResNet50 Ensemble with Grad-CAM]

Model Description

This model is an ensemble of three deep learning architectures: Vision Transformer (ViT), VGG19, and ResNet50. The ensemble approach enhances classification performance on medical image datasets related to ocular diseases. The model also integrates Grad-CAM visualization to highlight regions of interest for better interpretability.

Model Details

• Model Name: ROYXAI
• Developed by: Avishek Roy Sparsho
• Framework: PyTorch
• Ensemble Method: Bagging
• Backbone Models: Vision Transformer, VGG19, ResNet50
• Target Task: Medical Image Classification
• Supported Classes:
  • OT
  • Healthy
  • SC_diabetes
  • SC_cataract
  • SC_glucoma

Dataset

• Dataset Name: Custom Ocular Disease and its Secondary complications Dataset
• Dataset Source: Private Dataset (Medical Images)
• Dataset Structure: Images stored in folders based on class labels
• Preprocessing:
  • Resized images to 224x224 pixels
  • Normalized using ImageNet mean and standard deviation

Model Performance

• Accuracy: 98% on the test dataset
• Precision/Recall/F1-score: Evaluated and optimized for medical diagnosis
• Overfitting Prevention: Implemented data augmentation, dropout, weight regularization

Installation and Usage

Clone the Repository

git clone https://huggingface.co/Aviroy/ROYXAI
cd ROYXAI

Install Dependencies

pip install -r requirements.txt

Training the Model

To train the model from scratch, run:

python train.py --epochs 50 --batch_size 32

Load Pretrained Model

To directly use the trained model:

import torch
from PIL import Image
import torchvision.transforms as transforms
from model import ensemble_model  # Load the trained ensemble model

# Define image transformations
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# Load and preprocess an image
image_path = "path/to/image.jpg"
image = Image.open(image_path).convert('RGB')
image = transform(image).unsqueeze(0).to('cuda' if torch.cuda.is_available() else 'cpu')

# Perform inference
ensemble_model.eval()
with torch.no_grad():
    output = ensemble_model(image)
    predicted_class = torch.argmax(output, dim=1).item()

# Print classification result
print("Predicted Class:", predicted_class)

Grad-CAM Visualization

Visualizing Attention Maps for Interpretability

Vision Transformer (ViT)

from visualization import visualize_gradcam_vit  # Function for ViT Grad-CAM

# Generate Grad-CAM visualization
overlay = visualize_gradcam_vit(ensemble_model.models[0], image, target_class=predicted_class)

# Display the Grad-CAM output
import matplotlib.pyplot as plt
plt.imshow(overlay)
plt.axis('off')
plt.title("Grad-CAM for Vision Transformer")
plt.show()

ResNet50

from visualization import visualize_gradcam  # General Grad-CAM function

# Generate Grad-CAM visualization for ResNet50
overlay = visualize_gradcam(ensemble_model.models[2], image, target_class=predicted_class)

# Display the Grad-CAM output
import matplotlib.pyplot as plt
plt.imshow(overlay)
plt.axis('off')
plt.title("Grad-CAM for ResNet50")
plt.show()

VGG19

from visualization import visualize_gradcam  # General Grad-CAM function

# Generate Grad-CAM visualization for VGG19
overlay = visualize_gradcam(ensemble_model.models[1], image, target_class=predicted_class)

# Display the Grad-CAM output
import matplotlib.pyplot as plt
plt.imshow(overlay)
plt.axis('off')
plt.title("Grad-CAM for VGG19")
plt.show()

Training Configuration

• Optimizer: Adam with weight decay
• Learning Rate Scheduler: Cosine Annealing LR
• Loss Function: Cross-Entropy Loss
• Batch Size: 32
• Training Epochs: 20
• Hardware Used: T4 GPU x2 ,M1chip ,GPU P100

Limitations & Considerations

• This model is trained on a specific dataset and may not generalize well to other medical image datasets without fine-tuning.
• It is not a substitute for professional medical diagnosis.
• The Vision Transformer model is computationally expensive compared to CNNs.

Citation

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

@article{Sparsho2025,
  author    = {Avishek Roy Sparsho},
  title     = {ROYXAI Model For Proper Visualization of Classified Medical Image},
  journal   = {Medical AI Research},
  year      = {2025}
}

Acknowledgments

Special thanks to the open-source community and Kaggle for providing medical datasets for deep learning research.

License

This model is released under the Apache 2.0 License. Use it responsibly.