Multi-Disease-Aware Training Strategy for Cardiac MR Image Segmentation

TL;DR

The paper tackles the challenge of poor RV segmentation in cardiac MRI caused by distribution shifts across slices, phases, and disease conditions. It reframes this as a domain-generalization problem and introduces the Multi-Disease-Aware Training Strategy (MTS), which restructures data into disease-specific sets and employs Incomplete Training Data (ITD) via a rectangular ideal mask to improve generalization. A min-max learning objective over multiple diseases drives the training, and experiments on public ACDC and private DCWC datasets show substantial improvements in RV segmentation and cross-disease robustness, with ablations highlighting the importance of both MTS and ITD. The work demonstrates that incomplete, purposefully destroyed training data can enhance generalization, offering a principled approach to address distribution shifts in medical image segmentation and guiding future semi-supervised or unsupervised extensions.

Abstract

Accurate segmentation of the ventricles from cardiac magnetic resonance images (CMRIs) is crucial for enhancing the diagnosis and analysis of heart conditions. Deep learning-based segmentation methods have recently garnered significant attention due to their impressive performance. However, these segmentation methods are typically good at partitioning regularly shaped organs, such as the left ventricle (LV) and the myocardium (MYO), whereas they perform poorly on irregularly shaped organs, such as the right ventricle (RV). In this study, we argue that this limitation of segmentation models stems from their insufficient generalization ability to address the distribution shift of segmentation targets across slices, cardiac phases, and disease conditions. To overcome this issue, we present a Multi-Disease-Aware Training Strategy (MTS) and restructure the introduced CMRI datasets into multi-disease datasets. Additionally, we propose a specialized data processing technique for preprocessing input images to support the MTS. To validate the effectiveness of our method, we performed control group experiments and cross-validation tests. The experimental results show that (1) network models trained using our proposed strategy achieved superior segmentation performance, particularly in RV segmentation, and (2) these networks exhibited robust performance even when applied to data from unknown diseases.

Figures (10)

• Figure 1: Several examples of cardiac images from different slices. NOR and DCM represent normal subjects and dilated cardiomyopathy, and ED and ES denote end-diastolic and end-systolic phases. The blue, green, and red parts in the figures are the RV, MYO, and LV, respectively.
• Figure 2: A control group experimental illustration. The diagonal comparison is the main experiment, and the others belong to the ablation study.
• Figure 3: Box plots about segmentation evaluations of UNet on ACDC testing data. (a) NTS vs. MTS* when CTD, (b) NTS vs. MTS* when ITD, (c) CTD vs. ITD* when NTS, (d) CTD vs. ITD* when MTS. Notably, the former vs. the latter, where the former lacks a star *, while the latter includes one.
• Figure 4: Box plots about segmentation evaluations of TUNet on ACDC testing data. (a) NTS vs. MTS* when CTD, (b) NTS vs. MTS* when ITD, (c) CTD vs. ITD* when NTS, (d) CTD vs. ITD* when MTS. Notably, the former vs. the latter, where the former lacks a star *, while the latter includes one.
• Figure 5: Box plots about segmentation evaluations of UNet on DCWC testing data. (a) NTS vs. MTS* when CTD, (b) NTS vs. MTS* when ITD, (c) CTD vs. ITD* when NTS, (d) CTD vs. ITD* when MTS. Notably, the former vs. the latter, where the former lacks a star *, while the latter includes one.