Uncertainty-aware Mean Teacher Framework with Inception and Squeeze-and-Excitation Block for MICCAI FLARE22 Challenge

Hui Meng, Haochen Zhao, Ziniu Yu, Qingfeng Li, Jianwei Niu

Keywords: semi-supervised learning, multi-organ segmentation, uncertainty estimation, multi-scale features

Abstract:

Semi-supervised learning has attracted extensive attention in the field of medical image analysis. However, as a fundamental task,
semi-supervised segmentation has not been investigated sufficiently in the field of multi-organ segmentation from abdominal CT. Therefore, we propose a novel uncertainty-aware mean teacher framework with inception and squeeze-and-excitation block (UMT-ISE). Specifically, the
UMT-ISE consists of a teacher model and a student model, in which the student model learns from the teacher model by minimizing segmentation loss and consistency loss. Additionaly, we adopt an uncertainty-aware algorithm to make the student model learn accurate and reliable targets by making full use of uncertainty information. To capture multi-scale features, the inception and squeeze-and-excitation block are incoporated into the UMT-ISE. It is worth noting that abdominal CT of test cases are first extracted before multi-organ segmentation in the inference phase, which significantly improves segmentation accuracy. We implement experiments on the FLARE22 challenge. Our method achieves mean DSC of 0.7465 on 13 abdominal organ segmentation tasks.