Core Concepts
The core message of this article is to propose an Evidential Tri-Branch Consistency learning framework (ETC-Net) that employs three branches - an evidential conservative branch, an evidential progressive branch, and an evidential fusion branch - to effectively leverage both labeled and unlabeled data for semi-supervised medical image segmentation. The framework integrates evidential learning, uncertainty guidance, and evidential fusion to address critical issues such as prediction disagreement and label-noise suppression in cross-supervised training.
Abstract
The article introduces an Evidential Tri-Branch Consistency learning framework (ETC-Net) for semi-supervised medical image segmentation. ETC-Net consists of three branches:
Evidential Conservative Branch (ECB): This branch aims to generate cautious and conservative segmentation predictions, with fewer false positive regions.
Evidential Progressive Branch (EPB): This branch is designed to produce progressive and complete prediction results, minimizing false negative regions and complementing the ECB.
Evidential Fusion Branch (EFB): This branch leverages an evidence-based Dempster-Shafer fusion strategy to combine the predictions from ECB and EPB, generating more reliable and accurate pseudo-labels for unlabeled data.
The key highlights of the ETC-Net framework are:
It employs evidential learning to obtain predictions with uncertainty estimates, which are used to guide the cross-supervised training between ECB and EPB, mitigating the negative impact of erroneous supervision signals.
The bidirectional cross-supervised training between ECB and EPB enables the reliable exchange of complementary knowledge, addressing the confirmation bias issue in semi-supervised training.
The evidential fusion branch further reduces the noise in pseudo-labels, enhancing the efficiency of pseudo-label guided semi-supervised learning.
Extensive experiments on three medical image segmentation benchmarks (LA, Pancreas-CT, and ACDC) demonstrate that ETC-Net outperforms other state-of-the-art semi-supervised segmentation methods, significantly improving the segmentation performance compared to supervised training using only labeled data.
Stats
The LA dataset contains 100 gadolinium-enhanced MRI scans with an isotropic resolution of 0.625×0.625×0.625mm³.
The Pancreas-CT dataset includes 82 abdominal CT scans with a resolution of 512 × 512 and an isotropic resolution of 1 × 1 × 1mm³.
The ACDC dataset contains 200 annotated short-axis cardiac cine-MRI scans from 100 subjects.
Quotes
"To address this challenge, semi-supervised learning techniques have emerged, harnessing both labeled and unlabeled data to achieve segmentation performance comparable to fully supervised methods using exclusively labeled data."
"To handle the confirmation bias and error accumulation issues in semi-supervised medical image segmentation guided by pseudo labels, this paper proposes an evidence-based Tri-Branch consistency learning method."
"Benefiting from the complementary attributes of ECB and EPB and the evidence-based decision fusion strategy, the exploration and transfer of valuable unlabeled knowledge for segmentation improvement is further guaranteed."