MOTDNet: Multi organ task decoupling network for cell segmentation

In histopathology, cell detection and segmentation of Hematoxylin and Eosin (H&E) stained tissue images are essential clinical tasks. Existing methods decompose cell segmentation into three tasks: nucleus shape capture, overlapping nucleus separation, and nucleus type identification. However, these methods use a same structure decoder for all three tasks, ignoring the unique characteristics of each task and causing gradient conflicts, which ultimately affect segmentation accuracy. To address these issues, we introduce MOTDNet, a universal heterogeneous task decoupling network that incorporates directional pixel differential convolution(DPDC) and outer product convolution(OPC), along with task-specific decoders, enabling the segmentation and classification of 5 cell types across 19 tissues. Specifically, we design a CNN-SSM block to accurately capture cell shapes. Additionally, A directional differential convolution is designed to efficiently separate overlapping nuclei. Besides, we introduce an outer product convolution for accurately extracting contextual information in pathological images with numerous cell types and complex structures. Moreover, the outputs of the task-specific decoders are processed using the watershed algorithm to generate the final predictions. Extensive experimental reveal that our model not only achieves satisfactory performance on the PanNuke, MoNuSeg, Kumar and CPM17 medical datasets, but also attains competitive performance on the NYU-V2 and Cityscape non-medical datasets. Notably, our model achieves SOTA performance on the challenging PanNuke dataset with 13 GFLOPs of computation, 32M parameters, an F-1 score of 0.83, and a mean panoptic quality of 0.4999.