Heatmap-guided balanced multi-task learning approach for glistening characterization in OCT images

Abstract

[Abstract]: Glistening in intraocular lenses (IOLs) refers to the formation of water-filled microvacuoles within the lens material, typically assessed via slit-lamp (SL) photography. Recent proposals, like Optical Coherence Tomography (OCT) combined with deep learning, offer promising alternatives for evaluating glistenings. This study proposes a novel, automated multi-task method for detecting glistenings and segmenting the IOL area (IOLa) using heatmap-guided segmentation. Exhaustive experimentation demonstrates the method's generalization and stability across diverse IOL models with varying glistening severity, offering a dataset 375% more variable than previous studies. The study focuses on two tasks: glistening detection (Task I) and IOLa segmentation (Task II). For Task I, two single-task baselines were tested: a classical deep learning approach (a) and heatmap-guided segmentation (b). The best configurations achieved Intraclass Correlation Coefficient (ICC) scores of 0.901 for baseline (a) and 0.946 for baseline (b), while the multi-task approach reached 0.939. In Task II, improvements in segmentation accuracy were less pronounced but still achieved high Dice scores, consistently above 0.910. This research is the first to apply balanced multi-task strategies and heatmap-guided segmentation for glistening detection and to tackle IOLa segmentation. The method's competitive results confirm the value of deep learning for characterizing glistenings in OCT images, utilizing a more diverse and clinically representative dataset. This approach advances the field by providing a more reliable and efficient method for glistening evaluation.