Annotation-Efficient Deep Learning for Mediastinal Lymph Node Detection in CT

Overview

This study presents a semi-supervised deep learning pipeline that efficiently detects, segments, and measures mediastinal lymph nodes in contrast-enhanced CT scans. By combining expert annotations with pseudolabeled data and anatomical filtering, the method achieves performance comparable to fully supervised models while requiring significantly fewer manual annotations.

Background

Manual detection and measurement of mediastinal lymph nodes in CT imaging is essential for oncology staging and treatment planning but is time-consuming and prone to observer variability. Fully supervised deep learning models require large annotated datasets, which are often unavailable. Semi-supervised approaches leveraging both annotated and unannotated data can reduce annotation burden but face challenges with small structures like lymph nodes. Incorporating anatomical context and filtering can improve detection accuracy and reduce false positives.

Data Highlights

The proposed pipeline involves four key steps: (1) ensemble training of 3D nnU-Net models on limited expert-annotated scans; (2) generation of pseudolabels on unannotated scans by combining ensemble outputs; (3) anatomical filtering to remove false positives based on mediastinal structure constraints; and (4) final training of a single 3D nnU-Net model on filtered pseudolabels. This approach requires one-fourth to one-eighth the annotated data compared to fully supervised methods while maintaining similar performance.

Key Findings

• The semi-supervised method effectively combines expert annotations with pseudolabels to reduce the need for extensive manual labeling.
• Anatomical filtering significantly reduces false positive lymph node detections by leveraging mediastinal structure segmentations.
• The final 3D nnU-Net model trained on filtered pseudolabels achieves detection and segmentation performance comparable to fully supervised models.
• The approach addresses challenges of detecting small, clustered lymph nodes with fuzzy boundaries in volumetric CT scans.
• Observer variability in lymph node detection and measurement remains high, underscoring the need for automated, reliable methods.

Clinical Implications

This annotation-efficient pipeline can streamline the workflow of radiologists by automating mediastinal lymph node detection and measurement, reducing time and variability. The method's reduced dependency on large annotated datasets facilitates broader clinical adoption and supports more consistent oncology staging and treatment planning.

Conclusion

The presented semi-supervised deep learning approach offers a practical and reliable solution for mediastinal lymph node analysis in CT imaging, combining annotation efficiency with high accuracy. Its integration into clinical practice could enhance diagnostic consistency and efficiency.

References

1. Isensee et al. 2021 -- nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation
2. Totalsegmentator 2022 -- Automated segmentation of 104 anatomical structures in CT scans
3. Oda et al. 2022 -- 3D U-Net for lymph node detection with auxiliary anatomical labels
4. Bouget et al. 2021 -- Mask R-CNN and 2D U-Net for mediastinal lymph node detection and segmentation
5. Mathai et al. 2023 -- Ensemble 3D nnU-Net models for lymph node segmentation
6. McErlean et al. 2019 -- Observer variability in lymph node detection and measurement
7. Hopper et al. 2020 -- SAL measurement variability in lymph nodes
8. Fabel et al. 2021 -- Comparison of manual and computed lymph node measurements on CT