RetinaLiteNet: A Lightweight Transformer based CNN for Retinal Feature Segmentation

Mehwish Mehmood, Majed Alsharari, Shahzaib Iqbal, Ivor Spence, Muhammad Fahim; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 2454-2463

Retinal image analysis plays a pivotal role in diagnosing diseases like glaucoma diabetic retinopathy neurodegenerative disorders and cardiovascular diseases. The recent advancement of artificial intelligence (AI) can assist the practitioners to analyze the images accurately. In this research a lightweight deep learning model is proposed which is based on multitask learning to segment the retinal images including retinal vessels and optic disc for further analysis by clinicians. The proposed model has encoder-decoder framework where the encoder has convolutional layers with multi-head attention that captures both local details and long-range dependencies effectively. The resulting features from convolutional layers and multi-head attention are fused together to make the model more efficient and resilient for segmentation tasks. To further refine the features the skip connections are implemented along with the convolutional block attention module (CBAM) in the decoder. The model's efficiency is validated on two publicly available datasets (i.e. IOSTAR and DRIVE) to confirm the lightweight aspects and robustness. It achieved the dice scores of 80.6% and 93.3% on DRIVE and 80.1% and 85.4% on IOSTAR dataset for simultaneous segmentation of blood vessels and optic disc respectively. The empirical evaluations show 0.25 MB of memory 0.066 million parameters and a FLOPs estimation of 2.46 GFLOPs which is better than existing models.