Glaucoma Detection with MobileViT and cdr

I Implemented a Glaucoma detection model using the MobileViT architecture, trained on the ORIGA dataset.

Model Architecture

The model is based on the MobileViT architecture, Key components include:

• The model uses the pre-trained apple/mobilevit-small as its backbone
• unfreeze the final MobileViT encoder layer and its last convolutional layer for fine-tuning
• A global average pooling layer is applied to the feature maps outputted by MobileViT
• The classifier takes the pooled features and cdr (cup-to-disc ratio)
  • First layer: 640 + 1 input features (640 from MobileViT, 1 for cdr)
  • Second layer: 256 input features, 2 output features

Dataset

The ORIGA dataset is used for training and validation. It can be downloaded from Kaggle: https://www.kaggle.com/arnavjain1/glaucoma-datasets

Considerations

The model achieves 96% validation accuracy during training, However, this high accuracy may indicate overfitting to the validation data so consideration is needed