Modification of One-Stage Detector Architecture for Improving Small-Object Instance Segmentation in Images

For the first time, the problem of segmenting eight key optical coherence tomography biomarkers is formulated as an instance segmentation problem. To this end, a new data set has been created with manual annotation of instances for all eight types of biomarkers, providing a basis for comprehensive model training. YOLOv8 has been chosen as the base model. It is a modern single-stage object detector that ensures high computational efficiency and supports joint learning of localization, classification, and segmentation problems within a single architecture. However, direct application of YOLOv8 revealed limited accuracy for very small objects. Therefore, a modification of this architecture is proposed, aimed at improving the detection of small biomarkers while maintaining computational efficiency. Specifically, a bidirectional feature pyramid network is integrated into the YOLOv8 architecture, which significantly enhances the segmentation accuracy of small retinal structures.