Transfer Learning-Based Detection of Endometrial Cancer Lesion Regions on MRI Images

Accurate localization of the lesion region on magnetic resonance imaging (MRI) images of patients with endometrial cancer (EC) facilitates subsequent diagnosis by computeraided diagnostic systems or clinicians. MRI images of 294 patients with early-stage EC were retrospectively studied. Four classical object detection models (SSD, Faster R-CNN, CenterNet, YOLOv4) were trained by using the transfer learning approach. The detection performances of each model for the lesion region were compared and analyzed. The results show that the SSD model has the best detection performance in the test dataset with the AP of 99.52% and F1 score of 0.99 for detecting uterus and the AP of 96.12% and F1 score of 0.91 for detecting tumor at an IoU threshold of 0.5. The SSD model could be a useful tool for the implementation of a computer-aided diagnosis system for EC patients on MRI images.