Automatic diagnosis of cardiac magnetic resonance images based on semi-supervised learning

TL;DR

This work tackles automatic cardiac MRI segmentation and disease diagnosis with minimal labeled data. It introduces a semi-supervised segmentation network that uses a channel-prior CPCA attention module and complementary decoders to leverage unlabeled data, coupled with a dual-layer ensemble classifier for five-class heart-disease prediction. Key contributions include the CPCA-enhanced segmentation, complementary consistency learning, and a two-stage classifier that effectively distinguishes MINF and DCM, achieving high accuracy even with limited annotations. The approach yields accurate segmentation, reliable clinical indices such as volumes and EF, and robust disease predictions, offering practical utility in clinical workflows; overall, the method reduces labeling burden while maintaining high performance, as evidenced by results on the ACDC dataset with $EF=rac{V_{ED}-V_{ES}}{V_{ED}} imes100\\\ ext{%}$ and related indices.

Abstract

Cardiac magnetic resonance imaging (MRI) is a pivotal tool for assessing cardiac function. Precise segmentation of cardiac structures is imperative for accurate cardiac functional evaluation. This paper introduces a semi-supervised model for automatic segmentation of cardiac images and auxiliary diagnosis. By harnessing cardiac MRI images and necessitating only a small portion of annotated image data, the model achieves fully automated, high-precision segmentation of cardiac images, extraction of features, calculation of clinical indices, and prediction of diseases. The provided segmentation results, clinical indices, and prediction outcomes can aid physicians in diagnosis, thereby serving as auxiliary diagnostic tools. Experimental results showcase that this semi-supervised model for automatic segmentation of cardiac images and auxiliary diagnosis attains high accuracy in segmentation and correctness in prediction, demonstrating substantial practical guidance and application value.

Figures (7)

• Figure 1: A schematic depiction of the semi-supervised automatic segmentation and auxiliary diagnostic process is presented. CMR 3D images and partial labels act as inputs, processed through a semi-supervised segmentation network to generate segmentation results. Subsequently, these segmentation outcomes are transformed into various features and supplied to a disease prediction classifier, culminating in diagnostic outcomes.
• Figure 2: A schematic diagram of the semi-supervised cardiac image segmentation network structure is depicted. The network comprises a shared encoder with a CPCA module and three decoders. Among them, two decoders, formed by alternately utilizing skip connections, constitute complementary auxiliary decoders. The probability maps generated by the decoders are sharpened to obtain pseudo-labels for mutual learning between the decoders.
• Figure 3: Diagram of the dual-layer ensemble classifier structure.
• Figure 4: Visualization of segmentation results for comparison purposes. Blue represents the right ventricle (RV). White represents the left ventricular myocardium (MYO). Red represents the left ventricle (LV).
• Figure 5: Error analysis using Bland-Altman plot and linear regression, with semi-supervised training on 7 annotated data (10%).