Deep learning-based auto-segmentation and radiomics classification cascade model for endometrial cancer based on MRI: A Multi-center study

Motivation: Endometrial cancer's molecular subtype influences treatment and prognosis. Radiomic models using MR features enable non-invasive subtype prediction, but manual lesion delineation is time-consuming. Goal(s): Develop an automated segmentation and radiomics classification model for endometrial carcinoma molecular subtypes. Approach: Created an automatic segmentation model using T2WI, DWI, and T1WI+C images. Validated performance with an external set. Established a radiomic classification model and evluated automatic ROI segmentation. Results: The average DSC of T2WI, DWI, and T1WI+C were 76.2%, 85.8%, and 80.5%, respectively. For the radiomic classification model of MSS/MSI , the AUC values for manual and automatic segmentation were 0.757、0.744（RF）, and 0.27、0.719（SVM），demonstrating excellent classification performance. Impact: We have established an automated segmentation-radiomics classification cascade model for identifying molecular subtypes of endometrial cancer. This model could be used for assisting radiologists in screening the molecular subtypes of endometrial cancer, demonstrating its promising clinical application prospects.