Quantitative evaluation of myometrial infiltration depth ratio for early endometrial cancer based on deep learning

Abstract

The diagnosis of Endometrial Cancer (EC) was usually made empirically by radiologists, which introduces subjectivity and error. Computer-aided diagnosis methods, rely excessively on deep learning, which makes people difficult to understand the diagnostic process intuitively due to its black-box nature. This study combines deep learning and radiologist experience to address these issues. Firstly, we adopt the U-net to segment the tumor and uterus from the EC magnetic resonance imaging (MRI) images. Secondly, three myometrial infiltration (MI) depth assessment algorithms (fast thinning, fit-ellipse and area ratio) have been employed to deal with the segmentation images. Finally, A convolutional neural network (CNN) classifier has been trained to choose the best depth assessment algorithms according to the different shapes of uterus and tumor. The fit-ellipse MI depth assessment algorithm outperforms the other two algorithms (fast thinning and area ratio) at detecting the sub-stage of the EC, within stage I. Compared to using a single algorithm, the CNN classifier is able to select the optimum algorithm based on the shapes of the uterus and tumor, with an accuracy of 93.34%, for the test samples. In addition, the CNN classifier demonstrated a better performance in evaluating the trend of tumor spreading in early EC, which may help specialists to better evaluate the tumor spreading trend of early EC and make better decisions. It also could be a part of the EC monitoring process, for patients on medication or undergoing fractional curettage.