A Deep Learning Classification Method for Early Endometrial Cancer on MRI Images

Abstract

Early detection benefits patients with endometrial cancer (EC). In order to detect and diagnose early endometrial cancer from an image with FIGO stage I endometrial cancer from an image without endometrial cancer on magnetic resonant image, a deep learning endometrial cancer classifier based on deep learning was presented. Data augmentation was used in the classifier training to address the data imbalance problem. The experimental results show the model test accuracy for axial DWI images is 0.870 and the area under the curve (AUC) of the receiver operating characteristic (ROC) curve is 0.8611 in original data and 0.9764 in augmented data. The model test accuracy for axial T2WI images is 0.810 and the AUC of the ROC curve is 0.9342 in original data and 0.9389 in augmented data. The model test accuracy for sagittal T2WI images is 0.677 and the AUC of the ROC curve is 0.7919 in original data and 0.9294 in augmented data. The deep learning EC classifier is potential for use in the computer-aided diagnosis for endometrial cancer on MRI images.