Refining feature representation for accurate fundus lesion segmentation.

BACKGROUND

Accurate segmentation of retinal lesions is essential for early detection of age-related macular degeneration (AMD), particularly its key indicators, choroidal neovascularization (CNV) and choroidal non-perfusion (CNP). However, the highly variable shapes and appearances of lesions pose substantial challenges for automated segmentation.

PURPOSE: To tackle these challenges, we propose a deep-learning model, the Retinal Feature Enhancement Network (RFENet), designed to improve the accuracy and robustness of retinal lesion segmentation under challenging imaging conditions.

METHODS

RFENet builds upon the UNeXt backbone and introduces two modules tailored for retinal lesion segmentation: an Adaptive Feature Refinement Unit (AFRU), which selectively emphasizes informative features, and an Optimized Channel-Wise Convolution Unit (OCCU), which captures fine structural details in irregular lesions. Two expert-annotated datasets were used: 1070 OCT B-scans from 83 CNV patients and 184 FFA images from 107 CNP patients. Images were split at the patient level into training (70%), validation (10%), and test (20%) sets. Performance was compared against state-of-the-art segmentation models, including UNet++, Swin-UNet, SelfReg-UNet, DAEFormer, and Mamba-UNet. Evaluation metrics included Dice coefficient and Intersection over Union (IoU). Statistical significance was assessed using paired two-tailed t-tests ( α = 0.05 $\alpha = 0.05$ ), and effect sizes were reported with Cohen's d $d$ .

RESULTS

On the CNV dataset, RFENet achieved a Dice of 82.50% and an IoU of 71.66%, with clear improvements over competing models. On the CNP dataset, RFENet achieved a Dice of 74.43% and an IoU of 60.80%, slightly surpassing the best benchmark. Most pairwise comparisons were statistically significant, including CNV compared with Swin-UNet and SelfReg-UNet (p < $<$ 0.001) and CNP compared with DAEFormer (p = $=$ 0.035). Statistical analyses, including effect size evaluation, further confirmed the robustness of these improvements, with the most notable gains observed in the CNV dataset.

CONCLUSIONS

RFENet demonstrated consistent improvements over strong benchmarks on both CNV and CNP tasks, with statistical evidence supporting the reliability of these gains. These results indicate that RFENet provides a reliable technical advance for automated segmentation of CNV and CNP lesions associated with AMD. To facilitate reproducibility and further research, the source code is publicly available at https://github.com/HaoSun223/RFENet.