Learning Consistency- and Discrepancy-Context for 2D Organ Segmentation

Recently, CNN-based methods lead tremendous progress in segmenting abdominal organs (e.g., kidney, liver, and pancreas) and anomaly tumors in CT scans. Although 3D CNN-based methods can significantly improve accuracy by using 3D volume as input, they need more computational cost and may not satisfy the efficiency requirement for many practical applications. In this study, we mainly aim at improving the 2D segmentation by leveraging the consistency- and- discrepancy-context information from adjacent slices. Specifically, the consistency context mainly considers that the prediction variance of two adjacent slices needs to follow the variance in the ground truth. The discrepancy-context assumes the label difference of adjacent slices usually occurs in the edge area of organs. To fully utilize the above context information, we further devise a two-stage 2.5D segmentation framework based on the U-Net that takes three adjacent slices as input. In the first stage, we encourage the predictions of the three slices following the consistency context. In the second stage, we refine the segmentation result by adopting the prediction discrepancy area of adjacent slices as an extra input. Experimental results on several challenging datasets demonstrate the effectiveness of our proposed methods. Moreover, the adjacent-slice context information considered in this study can be effortlessly incorporated into other segmentation frameworks without extra testing overhead.

Automated summary

This study aims to improve 2D segmentation accuracy by leveraging consistency and discrepancy context information from adjacent slices, proposing a two-stage 2.5D segmentation framework based on U-Net. Experimental results demonstrate the effectiveness of the proposed methods in improving segmentation accuracy.