A novel MOSnet model for multi-object segmentation of medical images in micro-computed tomography.

Objectives: Micro-computed tomography (Micro-CT) is renowned for its high resolution, holding a pivotal role in advancing medical science research. However, compared to CT medical imaging datasets, there are fewer publicly available Micro-CT datasets, especially those annotated for multiple objects, leading to segmentation models with limited generalization abilities.

Methods: In order to improve the accuracy of multi-organ segmentation in Micro-CT, we developed a novel segmentation model called MOSnet which can utilize annotations from different datasets to enhance the whole segmentation performance. The proposed MOSnet includes a control module coupled with a reconstruction block that forms a multi-task structure, effectively addressing the absence of complete annotations.

Results: Experiments on 85 contrast-enhanced micro-CTscans and 140 native micro-CTscans for mice demonstrate that MOSnet is superior to the most of advanced segmentation networks. Compared to the best results of ResUnet, Unet3+, DAVnet3+ and AIMOS, our method improved dice similarity coefficient by 4.1 and 2.4 %, increased jaccard similarity coefficient by 4.1 and 3.1 %, and reduced HD95 by 16.3 and 19.3 % on the two datasets respectively at least.

Conclusions: Our proposed model proves to be a robust and effective method for multi-organ segmentation in micro-CT, especially in situations where comprehensive annotations are lacking within a dataset.