Clinically applicable semi-supervised learning framework for multiple organs at risk and tumor delineation in lung cancer brachytherapy

Purpose

The generalization ability of deep learning-based automatic segmentation techniques for lung cancer in practical clinical applications remains under-validated. We reported an investigation that validated a robust semi-supervised conditional nnU-Net (SSC-nnUNet) model in multiple organs at risk (OARs) and tumor segmentation in lung cancer brachytherapy, also explored its potential in robot-assisted puncture diagnosis and treatment.

Materials and methods

Six hundred seventy-four patients with CT data from four partially labeled datasets were divided into training and validation sets at a ratio of 4:1, 181 patients from multiple centers (private dataset) with fully annotated data provided by 3 experienced radiation experts were used for testing comparison. Six experienced experts from multiple centers were asked to correct model-generated contours, and 8 junior oncologists were assigned to delineate contours based on model supporting. To verify the feasibility of the contouring model in robot-assisted surgical operations, an equivalent human model experiment was designed specifically for lung cancer puncture treatment.

Results

In model-based experienced expert assessment, the mean revision degree achieved a competitive score of 1.38 % by 6 multicenter experts. In model-based junior oncologist assessment, they acquired a mean revision degree and efficiency improvement of −1.82 % and 83.4 %, respectively. Guided by the segmentation results of OARs and tumors, an average puncture error of 0.78 mm was achieved across 10 puncture experiments.

Conclusion

The SSC-nnUNet model showed a significant improvement in the segmentation quality and efficiency especially in junior oncologist delineation. Specifically, robot-assisted experiments illustrated that the model has great application potential in clinical treatment.