Early Detection of Retinal Microaneurysms Through Neurovascular Remodeling-Aware Transformer-LSTM Network

Abstract

Retinal microaneurysms (MAs), crucial for early diabetic retinopathy detection via fundus imaging, are difficult to identify due to the complexity of retinal pathology. Hence, a novel “Neurovascular Fourier-Otsu Vision Support Transformer Network with Long Short-Term Memory (NFOVSTN-LSTM)” is proposed to improve retinal MA detection and analysis using retinal fundus imaging. However, macular degeneration caused by neurovascular remodeling leads to vascular alterations and intensity fluctuations that are not directly associated with visible lesions. Thus, a novel Neurovascular Remodeling Detection Network (NRDN) is presented to enhance vascular contrast and texture by capturing subtle variations and detecting irregular branching, capillary dropout, and fragmented vascular patterns. In addition, extracting MA candidates is challenging due to the dense neural architecture of the region, which generates significant background clutter and variability in image texture. Hence, a Fourier-Otsu Hough Graph Transformer Network (FOHGTN) is used to improve MA localization by reducing background clutter, identifying subtle circular lesions, and capturing spatial and anatomical relationships. Meanwhile, silent ischemia complicates early detection of MAs, due to reduced perfusion in the absence of visible retinal biomarkers. Hence, a Local Vision Support Machine Transformer (LVSMT) is introduced to accurately distinguish MAs from ischemic regions by combining local texture features with global context, enabling precise classification even in the absence of visible biomarkers. Experimental results demonstrate improved accuracy, higher sensitivity, and lower loss, confirming the proposed method's superior performance in MA detection compared to existing approaches.