2026-10-14T10:41:22Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32805629 2026-10-01T16:14:31Z portal_1172 item_type_8 month_year_10_2026

Lightweight deep learning-based retinal fundus image segmentation Mehwish Mehmood (8378757) PUREID: 683962963 Lightweight Deep Learning Models Retinal Image Analysis Fundus Images Image Segmentation Retinal image analysis is crucial to identify many ocular, systemic, and neurodegenerative diseases. However, precise and efficient retinal feature segmentation remains challenging due to the complex and low-contrast vascular structures. Existing deep learning models used to segment retinal vessels with high accuracy require extensive memory and computational resources. This thesis addresses these limitations by creating a series of lightweight, high-performance deep learning architectures for retinal feature segmentation in resource-constrained and real-time clinical settings. The research presents four progressively optimized models that demonstrate a systematic progression from hybrid Transformer-CNN designs to optimized, attention-based models, influenced by the latest Vision Mamba-based feature-mixing approaches. <br><br>The first model, RetinaLiteNet, uses a hybrid transformer framework to perform simultaneous segmentation of retinal blood vessels and the optic disc. By integrating convolutional feature extraction with multi-head attention and convolutional block attention in the skip connections, the model captures both local textures and global dependencies while being lightweight with only 0.066 million parameters. However, despite its strong performance, hidden computational overhead from transformer operations encouraged the development of architectures offering more predictable runtime efficiency. To address this, a hybrid model named LVS-Net is presented that adopts a fully convolutional encoder-decoder, which integrates focal modulation and spatial feature refinement. This architecture improves fine-grained vessel continuity, increases sensitivity to thin structures, and removes the transformer's implicit FLOP overhead with only 0.71 million parameters. Despite improved segmentation efficiency of LVS-Net, its attention mechanisms are less sensitive to regional differences in retinal contrast. Based on this limitation, the third lightweight model, LFRA-Net, introduces a compact network that combines focal modulation with region-aware attention at the bottleneck and also routes region awareness through selective skip connections. This design improves contextual understanding, improves the segmentation of thin and low-contrast vasculature, and reduces the parameter count to only 0.17 million. However, its sequential attention pipeline still constrains interaction between global and local contexts. To improve the performance of the models mentioned above, LFA-Net is proposed, which incorporates a multi-level LiteFusion-Attention mechanism that performs dynamic feature mixing, modulation-based attention, and residual learning within a unified module. This model captures both long-range dependencies and thin structures while remaining computationally efficient with only 0.11 million parameters. The models are evaluated across five publicly available datasets, including RITE, DRIVE, IOSTAR, STARE, and CHASE_DB, showing that they are suitable for point-of-care and mobile screening applications. <br><br>Overall, this thesis presents a coherent progression of lightweight segmentation models that strike a balance between accuracy, computational predictability, and efficiency. The model's evaluations of a hybrid Transformer-CNN-based approach, adaptable convolutional modulation, region-aware attention, and multi-level fusion indicate the path to deployable, real-time retinal image analysis systems. Eventually, these evaluations improve automated retinal feature segmentation, which facilitates the early detection of ocular, vascular, and neurodegenerative diseases. 2026-10-01T16:14:31Z Text Thesis 10.17034/32805629.v1 https://figshare.com/articles/thesis/Lightweight_deep_learning-based_retinal_fundus_image_segmentation/32805629 All Rights Reserved