Annotation-efficient deep learning detection and measurement of mediastinal lymph nodes in CT

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Manual measurement of mediastinal lymph nodes (LNs) in contrast-enhanced CT (ceCT) scans is standard but time-consuming and prone to variability.
• Deep learning models offer automated solutions but typically require extensive expert annotations, posing a significant bottleneck.

Purpose of the Study:

• To develop an annotation-efficient semi-supervised deep learning method for automatic detection, segmentation, and measurement of mediastinal LN short axis length (SAL) in ceCT scans.
• To reduce the manual annotation effort required for training AI models in medical imaging.

Main Methods:

• A semi-supervised approach combining expert annotations with pseudolabeled data from unannotated scans.
• Utilized an ensemble of 3D nnU-Net models to generate pseudolabels, followed by anatomical filtering to remove false positives.
• Trained a final 3D nnU-Net model on the filtered pseudolabels, optimizing the ratio of annotated to unannotated data.

Main Results:

• Semi-supervised models achieved recall improvements of 11-24% (0.72-0.87) compared to fully supervised methods, with comparable precision.
• The best model demonstrated mean SAL differences of 1.65 ± 0.92 mm for normal LNs and 4.25 ± 4.98 mm for enlarged LNs, falling within observer variability.
• Evaluated on three chest ceCT datasets comprising 268 annotated and 710 unannotated scans.

Conclusions:

• The proposed semi-supervised method requires significantly fewer annotations (one-fourth to one-eighth) to achieve performance comparable to fully supervised models for mediastinal LN measurement.
• Pseudolabeling combined with anatomical filtering presents a viable strategy to address annotation challenges in developing AI solutions for radiology.