Multi-label classification allows an image to receive more than one disease label, rather than forcing a choice between mutually exclusive categories. This project focuses on diabetic retinopathy (DR), glaucoma and age-related macular degeneration (AMD), using pretrained convolutional neural networks as the starting point for retinal-image classification.
The methods address several parts of that learning problem: Focal and class-balanced (CB) loss address unequal class representation, attention modules help models emphasise informative features, and ensembles combine predictions from multiple models. Test-time augmentation (TTA) combines predictions across transformed versions of an image.
Built a multi-label retinal imaging pipeline and fine-tuned CNNs for DR, Glaucoma, and AMD.
Developed the pipeline and contributed to model fine-tuning.
Used Python, PyTorch, ResNet18, EfficientNet, and pretrained CNNs.
Built a multi label medical imaging pipeline for DR, Glaucoma, and AMD.
Fine tuned pretrained CNNs.
Handled class imbalance with Focal and CB Loss.
Applied attention modules and ensemble strategies with TTA.
To discuss transfer learning and multi-label retinal image classification, get in touch.