DaCSeg: Divide and conquer for accurate overlapping chromosome instance segmentation in metaphase cell images

Xinyu Fan, Hui Liu, Hongdang Zheng, Jingfang Zhai, Lin Zhang

Abstract: Highlights•To obtain a well-built overlapping chromosome dataset, we propose a universal algorithm that automatically marks the overlapping region of chromosomes, named Overlapping Chromosome Annotation (OCA), which can be extended to other instance segmentation tasks.•To divide and conquer the overlapping chromosomes, we build an overlapping estimation branch to detect whether chromosome overlap occurs and then design the Overlapping Chromosomes Identification Module (OCIM) for detection. Next, we propose the Overlap-non-overlap Attention Segmentation Module (OASM) to unravel and explore the overlapping relationship between chromosomes and better capture the featrues of highly overlapping regions for mask prediction.•To further correct the segmentation prediction, we introduce the Shape Prior guided Segmentation Module (SPSM), which captures chromosome generic template shapes based on the behavior of human perception to train the model behavior.