SRD U-Net: Spatial Refined Depthwise U-Net forRetinal Vessel Segmentation

Abstract Precise pixel level categorization of retinal vessels requires both deep contextualawareness and distortion free spatial representation. Traditional encoder-decodermodels often suffer from excessive parameter counts and an inability to prioritizesignificant feature spaces selectively. Consequently sub optimal boundaries andnoise propagation are presents in the retinal vascular fundus images. To addressthese limitations, this paper introduces the contrast-aware retinal preprocessorblock(CARP) and spatial refined depthwise U-Net (SRD U-Net). The CARPserves as a retinal image preprocess and augmentation module that enhancescontrast and sharpness. The CARP uses contrast-limited adaptive histogramequalization and TensorFlow-based luma?chroma sharpening to generate preprocesseddatasets for segmentation tasks. The SRD U-Net is a lightweight highlydiscriminative segmentation network with three blocks including conventionalfilter depthwise unit (CFDU) encoder, channel attention bottleneck(CAB) andCFDU decoder. . The core of our proposed architecture lies in a novel hybridfeature-extraction sequence, CFDU. The CFDU block initially captures primitivespatial patterns through a standard 2D convolution. The convolution outputtransitioning to a manual depthwise separable configuration, decoupling spatialfiltering from channel mixing to dramatically compress the network parameter.Furthermore, we combat feature redundancy in the latent space by integratinga channel attention bottleneck(CAB). The CAB block aggregates global contextthrough spatial pooling to dynamically compute non-linear inter channel dependencies.The bottleneck attention mechanism explicitly recalibrates channel-wiseresponses with highly informative feature maps. In addition, the bottlenecksystematically suppresses attention, filtering out irrelevant background noise.Furthermore, we evaluated our framework with DRIVE, HRF, STARE, CHASEand FIVES datasets. The performance metrics used for assessment of proposedframework are accuracy, sensitivity, F1 score, recall and precision. On averageacross all datasets, our proposed method afforded improvements of approximately2 % among all metrics compared to the lowest-performing baseline. Finally,our model achieves comparable accuracy with significantly lower computationaloverhead, featuring reduced parameter size and faster execution speeds. Thisportrays our approach as a highly efficient alternative for resource-constrainedenvironments.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11327124/v1
Primary Topic
Retinal Imaging and Analysis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

SRD U-Net: Spatial Refined Depthwise U-Net forRetinal Vessel Segmentation

Nagaraj Yamanakkanavar, Srinivas Rangu, Narasimhadhan A V
Research Square
Retinal Imaging and Analysis
preprint

SRD U-Net: Spatial Refined Depthwise U-Net forRetinal Vessel Segmentation

Nagaraj Yamanakkanavar, Srinivas Rangu, Narasimhadhan A V
preprint en

Abstract

Abstract Precise pixel level categorization of retinal vessels requires both deep contextualawareness and distortion free spatial representation. Traditional encoder-decodermodels often suffer from excessive parameter counts and an inability to prioritizesignificant feature spaces selectively. Consequently sub optimal boundaries andnoise propagation are presents in the retinal vascular fundus images. To addressthese limitations, this paper introduces the contrast-aware retinal preprocessorblock(CARP) and spatial refined depthwise U-Net (SRD U-Net). The CARPserves as a retinal image preprocess and augmentation module that enhancescontrast and sharpness. The CARP uses contrast-limited adaptive histogramequalization and TensorFlow-based luma?chroma sharpening to generate preprocesseddatasets for segmentation tasks. The SRD U-Net is a lightweight highlydiscriminative segmentation network with three blocks including conventionalfilter depthwise unit (CFDU) encoder, channel attention bottleneck(CAB) andCFDU decoder. . The core of our proposed architecture lies in a novel hybridfeature-extraction sequence, CFDU. The CFDU block initially captures primitivespatial patterns through a standard 2D convolution. The convolution outputtransitioning to a manual depthwise separable configuration, decoupling spatialfiltering from channel mixing to dramatically compress the network parameter.Furthermore, we combat feature redundancy in the latent space by integratinga channel attention bottleneck(CAB). The CAB block aggregates global contextthrough spatial pooling to dynamically compute non-linear inter channel dependencies.The bottleneck attention mechanism explicitly recalibrates channel-wiseresponses with highly informative feature maps. In addition, the bottlenecksystematically suppresses attention, filtering out irrelevant background noise.Furthermore, we evaluated our framework with DRIVE, HRF, STARE, CHASEand FIVES datasets. The performance metrics used for assessment of proposedframework are accuracy, sensitivity, F1 score, recall and precision. On averageacross all datasets, our proposed method afforded improvements of approximately2 % among all metrics compared to the lowest-performing baseline. Finally,our model achieves comparable accuracy with significantly lower computationaloverhead, featuring reduced parameter size and faster execution speeds. Thisportrays our approach as a highly efficient alternative for resource-constrainedenvironments.

Research Square
Central University of Karnataka (IN), National Institute of Technology Karnataka (IN)
Retinal Imaging and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon