Fully automatic lesion segmentation and deep myometrial invasion prediction method for endometrial cancer based on MRI

Motivation: The judgment of myometrial invasion depth by traditional imaging is subjectively affected by observers. Manual segmentation is labor-intensive and not feasible in daily work. Goal(s): This study explores the use of nnU-Net for automatic segmentation of endometrial cancer, and uses radiomics to predict endometrial cancer deep myometrial invasion. Approach: 127 endometrial cancer patients were assigned to two cohorts in a 7:3 ratio. 50 cases of data from the training group were selected to establish a nnU-Net segmentation model. Results: Dice score of nnU-Net were 0.956 and 0.922 in the training and validation group. The AUCs of combined model were 0.903 and 0.894 respectively. Impact: NnU-Net has the potential to automatically identify and segment endometrial cancer lesions. The combined model integrating radiomics features and clinical risk factors has a better ability to identify deep myometrial invasion.