An Automatic Staging Method for Early Cervical Cancer on MRI Images: A Pilot Study

Abstract

Objective: To automatically segment sagittal mag-netic resonance imaging(MRI) images of cervical cancer (CC) patients and calculate the Intersection over Union (IoU) value of segmentation results for staging CC. Methods: This work uses 51 stage IB MRI images and 27 stage IIA MRI images as a dataset, all of which were pathologically confirmed in patients with CC. All labels used for training were manually labeled by radiologists with many years of clinical experience. MRI images and corresponding labels are put into a semantic segmentation network (based on U-net) to segment the tumor, uterus and vagina, respectively. Finally, the IoU values between the tumor and the other two tissues are calculated and the appropriate threshold is selected to classify images. Results: The segmentation of the uterus has good performance with Intersection over Union (IoU) of 0.872 and Pix Accuracy(PA) of 0.923. The classifier uses the best diagnostic threshold for tumor and vagina as the main criteria and the best diagnostic threshold for tumor and uterus as an auxiliary judgment to obtain a final classification accuracy of 0.807, the precision of 0.846, and recall of 0.863. Conclusions: In this study, the segmentation results were used to calculate the IoU values between different tissues which can distinguish the early staging of CC. It can provide some reference for radiologists to determine the tumor staging on MRI images of CC patients.