6533b7d0fe1ef96bd125b413

RESEARCH PRODUCT

Computer-aided-diagnosis for ocular abnormalities from a single color fundus photography with deep learning

subject

description

Any damage to the retina can lead to severe consequences like blindness. This visual impairment is preventable by early detection of ocular abnormalities. Computer-aided diagnosis (CAD) for ocular abnormalities is built by analyzing retinal imaging modalities, for instance, Color Fundus Photography (CFP). The main objectives of this thesis are to build two CAD models, one to detect the microaneurysms (MAs), the first visible symptom of diabetic retinopathy, and the other for multi-label detection of 28 ocular abnormalities consisting of frequent and rare abnormalities from a single CFP by using deep learning-based approaches. Two methods were proposed for MAs detection: ensemble-based and cascade-based methods. Ensemble-based MAs detection aims to find the best combination of input channels while the goal of cascade-based MAs detection is to reduce the false positive predictions with high sensitivity. The MAs detection with the cascade learning method achieves 0.792 sensitivity, the highest sensitivity on the E-Ophta dataset in 8 false positives per image. Two methods were also proposed for multi-label detection: Convolutional Neural Network (CNN)-based and Transformer-based methods. These proposed methods combine the visual features extracted from a color fundus image and the label co-occurrence dependencies extracted from linguistic features. The correlation between visual and linguistic features is learned by a semantic dictionary. CNN-based multi-label detection aims to adapt the model with out-of-vocabulary words. The results of this model show the positive impact of linguistic input interference in multi-label detection. Transformer-based multi-label detection enhances the linguistic input interference in multi-label detection. This method achieves a 0.804 final score, the highest score on the RFMiD Test dataset.