A Deep Learning-based High Precision Segmentation Method for Endometrial Cancer using MRI Images

Endometrial Cancer(EC) is among the most prevalent malignant tumors affecting the female reproductive system. Early treatment improves survival. The manual segmentation of the uterus and tumor by the physician on the Magnetic Resonance Image(MRI) is cumbersome time-consuming and susceptible to the subjective influence of the physician. Deep Learning(DL) technology has achieved satisfactory results in recent years, which can help doctors complete burdensome segmentation tasks. In this study, we investigate a novel segmentation model named Triplet Attention SegFormer(TASeg) that utilizes a three-branch structure to capture interactions between dimensions and compute attention weights. Triplet Attention establishes dimensional dependencies between input tensors followed by residual transformations and encodes inter-channel spatial information. This improves the segmentation performance of our model on EC datasets. The experimental results showed that the Mean Intersection over Union(MIoU) of three different sequences were $90.75,87.20$ and 83.94, respectively, superior to the existing segmentation methods.