Retinal neovascularization detection from fundus images using deformable attention ResNet and swarm-hiking optimization.

Abstract

Background: In the detection of retinal neovascularization (RNV) from fundus images, defining anatomically consistent perivascular regions is challenging due to asymmetric retinal tissue deformation caused by abnormal vessel proliferation in RNV, leading to inaccurate extraction of pathological features. Moreover, early pathological changes in perivascular zones are minute and blend with normal tissue variability, while existing models focused on global retinal features tend to miss these localized abnormalities, resulting in reduced detection accuracy.

Methods: A novel deep learning (DL) framework, Deformable Attention Pyramid ResNet with Perivascular Swarm-Hiking Optimization (DAPR-PSHO) is proposed for precise RNV detection. In this, a Deformable Attention Pyramid Scene Parsing-ResNet (DAPSP-ResNet) is employed, which adaptively adjusts its receptive fields using deformable convolutions to accurately capture irregular vessel structures, and enriches feature extraction by integrating local vessel details with global retinal context. Then, a PeriV-Transformer driven Particle Swarm-Hiking Optimization is introduced, to focus attention on minute perivascular changes while maintaining global retinal context, and fine-tunes transformer parameters, thereby effectively capturing early-stage RNV and enhancing anatomical fidelity indetection.

Result: The proposed RNV detection model attains a high accuracy of 0.98, precision of 0.985, and recall of 0.986 comparedto the existing models.

Conclusion: The DAPR-PSHO framework successfully addresses the limitations of global-feature-focused models by eff ectively isolating localized, early-stage abnormalities in perivascular zones, significantly improving RNV detection accuracy and anatomical fidelity.