A Lightweight UNet with Inverted Residual Blocks for Diabetic Retinopathy Lesion Segmentation

Amit Bhati, Karan Choudhary, Samir Jain, Neha Gour, Pritee Khanna, Aparajita Ojha, Naoufel Werghi

Abstract: Diabetic Retinopathy (DR) is a progressive disease that significantly contributes to vision impairment and blindness. Its complex nature, characterized by subtle variations among different grades and the presence of numerous important small features, poses a considerable challenge for accurate recognition. Currently, the process of identifying DR relies heavily on the expertise of physicians, making it a time-consuming and labor-intensive task. However, automated detection of specific lesions plays a crucial role in visualizing, characterizing, and determining the severity of DR. Timely detection of DR in its early stages is vital for diagnosis and can potentially prevent blindness through appropriate treatment. Nonetheless, segmenting lesions in fundus imaging is a challenging task due to variations in lesion sizes, shapes, similarities, and limited contrast with other parts of the eye, leading to ambiguous results. In this work, a shallow UNet-based architecture with inverted residual skip connections is proposed to segment lesion parts of DR disease. Performance of the model is evaluated on Indian Diabetic Retinopathy Image Dataset (IDRiD) and DDR datasets. Results show that the proposed model is able to distinguish different kinds of DR lesion parts with a very less number of parameters (3.3 M).