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

Abstract

Chromosome segmentation is essential for karyotyping analysis, which provides a reliable basis for subsequent disease diagnosis. Recent deep segmentation techniques can significantly boost the efficiency of chromosome micrographs processing. However, existing classical Mask R-CNN-based methods cannot deal with chromosome overlapping issues highly susceptible to appearing in clinical data. In our opinion, it is difficult for current computational methods to achieve accurate segmentation without a specific design for complex overlapping. Therefore, we propose a divide-and-conquer automatic framework, which shows better robustness for complex chromosome instance segmentation. Exploiting a series of specially designed modules, our end-to-end model first performs an estimation of whether the chromosomes overlap and then generates precise attention masks for overlapping areas. With the assistance of the shape prior knowledge, it produces final instance segmentation predictions for the chromosomes. The experimental results show that the proposed method achieves mean Average Precision (mAP) of 84.10%, AP50 of 98.61%, AP75 of 97.62%, Precision of 89.32%, Recall of 88.87%, and F1-score of 89.09%, which are superior to other state-of-the-art methods. Code and models are available at https://github.com/labiip/DaCSeg.