A semi-supervised multi-connection contrastive learning framework for x-ray lung segmentation based on mutual...

Deep learning techniques have demonstrated impressive results in medical image segmentation tasks. However, building labeled data for fully supervised models requires significant labor, which can be costly. Additionally, medical devices often prioritize higher data security, smaller size, and improved portability, which typically leads to limited computing and storage capabilities and necessitates offline model deployment.

Given these challenges, it is meaningful to conduct in-depth research on high-performance tiny offline models suitable for edge deployment. This study aims to pursue a higher-performance segmentation model while keeping the inference model small enough to be used in clinical practice where computing resources are usually limited.

This study proposes a semi-supervised framework based on contrastive learning for developing an organ contour segmentation model using a few labels. The framework employs multiple consistency alignment and mutual distillation mechanisms, in which its backbone can be adapted based on performance or speed requirements.

The framework was tested on three chest X-ray datasets with 128×128 resolution (JSRT, Montgomery County, and Shenzhen Hospital). When only using two labeled images, the Dice scores of lung segmentation were 0.9636, 95% CI [0.9633, 0.9640], 0.9596, 95% [0.9589, 0.9604], and 0.9527, 95% [0.9508, 0.9546], respectively, and the inference model parameters obtained were only 1.15 M.

These results indicate that the framework performance and potential for edge deployment are leading all similar studies, proving its suitability for clinical applications. (Code Address: https://github.com/mozixr/SemiMCD/tree/main).