A Feature Pyramid Fusion Network Based on Dynamic Perception Transformer for Retinal OCT Biomarker Image Segmentation

OCT biomarkers are important for assessing the developmental stages of retinal diseases. However, the biomarkers show diverse and irregular features in OCT images and do not have a fixed location. In addition, irregular and unevenly distributed biomarkers can be accompanied by damage to the retinal layers, which can lead to blurred boundaries, thus making it difficult for physicians to make judgments about biomarkers. Therefore, we propose a dynamic perception Transformer-based feature pyramid fusion network for segmenting retinal OCT biomarker images. Our network consists of two modules: the Feature Pyramid Fusion Module (FPFM) and the Dynamic Scale Transformer Module (DSTM). The FPFM connects features at different scales while incorporates an attention mechanism to emphasize the important features. The DSTM dynamically adjusts the scale of the fused features and captures their long-range dependencies, which enables us to preserve small biomarkers and adapt to complex shapes. In this way, we can cope with the challenge of excessive biomarker scale changes from a variable scale perspective. Our proposed model demonstrates good performance on a local dataset.