Simultaneous multiclass retinal lesion segmentation using fully automated RILBP-YNet in diabetic retinopathy

• Published in: Biomedical signal processing and control 2023-09, Vol.86, p.105205, Article 105205
• Main Authors: Geetha Pavani, P., Biswal, B., Gandhi, Tapan Kumar
• Format: Article
• Language: English
• Subjects: Diabetic Retinopathy; IDRiD dataset; Image Segmentation; Local Binary Patterns; Neural Networks; Parallel encoder
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2023.105205

•The proposed model, RILBP-YNet (Residual Inception Local Binary Pattern - YNetwork) is implemented for segmenting multiple lesions like Microaneurysms, Haemorrhages, Hard and Soft exudates.•The novelty of the proposed work lies in adding a parallel encoder to traditional encoder-decoder architecture to induce textural features to improve the segmentation results.•Both residual and inception modules are integrated into the network to segment the lesions which are irregular in shape and size.•The proposed model is trained and tested on the Indian Diabetic Retinopathy Image Dataset (IDRiD), and e-ophtha datasets and achieved competitive results compared to the existing methods in terms of AUC score. The proposed method is also tested on datasets like DRIVE, STARE, CHASE-DB1, and one locally real-time dataset. This paper presents a novel method for segmenting the multiple lesions like Microaneurysms, Haemorrhages, Hard and Soft exudates that occurs on the surface of the retina due to diabetes. If these lesions are left untreated in their early stage, they may cause irrevocable vision loss to diabetic patients. However, the manual assessment of these vision threatening diseases is prone to errors and time-consuming process. As a result, the proposed model, RILBP-YNet (Residual Inception Local Binary Pattern - YNetwork) is implemented as a dual encoder and single decoder architecture in which one encoder extracts the features from the fundus image. Whereas the parallel encoder extracts the textural features from Local Binary Patterns (LBPs) of the corresponding fundus image simultaneously. The extracted features of the two encoders are concatenated at the end of every encoder level and fed as input to the next encoder. Finally, the concatenated features of the last encoder are transposed by performing the up-sampling operation and fed as input to the first decoder. Furthermore, at the end of every decoder, the features of the decoder and concatenated features of the respective encoder are fused and provided as input to the next decoder. As a result, at the end of the last decoder, segmented masks of each lesion are generated. The novelty of the proposed work lies in adding a parallel encoder to traditional encoder-decoder architecture to induce textural features to improve the segmentation results. The proposed model is trained and tested on the Indian Diabetic Retinopathy Image Dataset (IDRiD), and e-ophtha datasets and achieved competitive results compared to