Semi-supervised deep learning approach for left ventricle segmentation with limited annotations

Abstract Background Accurate left ventricle (LV) segmentation from cardiac magnetic resonance (CMR) images is pivotal for diagnosing and managing cardiovascular disease. Current manual methods, though considered gold standards, suffer from high inter-observer variability and extensive time requirements (45–60 minutes/scan). Current automated tools, predominantly optimised for static end-diastolic and end-systolic frames, and often fail during mid-cycle phases, missing subtle wall motion abnormalities critical for early cardiomyopathy detection. Purpose This study aimed to develop a deep learning model that automates LV segmentation across the entire cardiac cycle, particularly addressing the challenge of limited availability of CMR images with segmented LV labels. Methods Our semi-supervised learning model employed the "Inherent Consistent Learning for Accurate Semi-supervised Medical Image Segmentation" method, as outlined in the recent literature. This advanced approach utilises both labeled and unlabeled data within a convolutional neural network (CNN) framework. Our model specifically designed to harness the spatial and temporal dynamics of the LV. The model was trained using the DAPA-LVH dataset, which includes data from 66 patients with type 2 diabetes and left ventricular hypertrophy, featuring comprehensive annotations for 10 patients across 25 cardiac phases per cycle, totaling 1,994 image slices. The data were splitted for training (64 patients), validation and testing (1 patient each). Performance was rigorously validated using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95), ensuring the model’s precision and reliability in clinical scenarios. Results Training dynamics revealed rapid improvement, with the loss significantly decreasing and stabilising at a low level, and a consistent DSC indicating reliable segmentation accuracy. The mean HD95 showed a sharp reduction, indicating enhanced spatial accuracy. Validation metrics further supported these findings, with an average DSC of 0.876 and a standard deviation of 0.260, demonstrating the model's overall precision. However, the variability indicated by the standard deviation highlights areas for improvement, particularly in handling complex cardiac motions. Additionally, the model processed each scan in approximately 3 minutes. This demonstrates a significant efficiency improvement in processing time compared to manual methods. Conclusion The developed deep learning model significantly improves the automation and accuracy of LV segmentation in CMR images. It demonstrates potential to enhance clinical workflows by reducing analysis time and increasing diagnostic consistency. Future efforts will focus on refining the model to address the challenges observed in specific cardiac phases and expanding its application to other aspects of cardiac imaging.Segmentation results Model Training Metrics