Objective:

To develop a semi-supervised deep learning method for the automatic detection, segmentation, and measurement of mediastinal lymph nodes in contrast-enhanced CT scans, enhancing clinical efficiency.

Key Findings:

• The method requires one-fourth to one-eighth less annotated data compared to fully supervised methods, making it more feasible for clinical application.
• Incorporation of anatomical context through structure-based filtering reduces false positives, enhancing reliability.
• The approach combines expert annotations with pseudolabels for improved detection and measurement, potentially increasing diagnostic confidence.

Interpretation:

The proposed semi-supervised method enhances the efficiency and accuracy of lymph node detection and measurement in CT scans, addressing the limitations of manual methods and previous automated approaches.

Limitations:

• Performance may vary based on the quality of initial expert annotations; future work should focus on improving annotation quality.
• The method's effectiveness on different anatomical regions or imaging modalities is not evaluated; further studies are needed to generalize findings.

Conclusion:

The study presents a novel, efficient approach for lymph node analysis in CT scans, potentially improving clinical workflows and outcomes in oncology, particularly in reducing observer variability.