The method consists of a 3D nnU-Net voxel classification model for LN detection (Fig. 1). It is a semi-supervised method combines the precision of expert annotations with the quantity advantages of pseudo-labeled data. It uses an ensemble of 3D nnU-Net models trained on a few expert-annotated scans to generate pseudo-labels on a large dataset of unannotated scans. The pseudo-labels are then filtered to remove false positive LNs by excluding LNs outside the mediastinum and LNs overlapping with other anatomical structures. Finally, a single 3D nnU-Net model is trained using the filtered pseudo-labels. Our method helps optimize the ratio of annotated/non-annotated dataset sizes to achieve the desired performance, thus reducing the manual annotation effort.

It was initially trained on a few annotated scans and used to generate LN pseudolabels for a larger dataset of unannotated scans. The LN pseudolabels were filtered with automatic segmentations of 20 non-LN mediastinal structures. An enhanced 3D nnU-Net model was then trained with the labeled and the selected pseudolabels. LNs with axial short axis >10mm were identified as enlarged. 

We obtained 298 chest CECT studies of patients with suspected mediastinal lymphadenopathy from three sources: 108 from our hospital, 100 from the Lnq2023 Challenge dataset, and 90 from an NIH dataset. They included 2,014 annotated LNs, 1,078 normal (short axis of 5-10mm) and 836 enlarged (> 10mm). The training/test set partition was 134/164 scans (1,069/945 LNs). An additional 710 unannotated scans were collected from our hospital for pseudolabel generation. The ratio of annotated/unnanoted scans used for training was 1 to 4.3. Detected LNs and their computed short axes were compared to the annotated ground truth.