3643 - Autoregressive Tracking Transformer for Whole-Body 3D Lymph Node Detection in CT Scans

Purpose/Objective(s): Identifying scatteredly-distributed and low-contrast LNs in 3D CT scans is highly challenging, even for experienced clinicians. Previous lesion and LN detection methods demonstrate effectiveness of 2.5D approaches (i.e, using 2D network with multi-slice inputs) by leveraging pretrained 2D model weights, and show improved accuracy as compared to 2D or 3D detectors. However, slice-based 2.5D detectors do not explicitly model inter-slice consistency for LN as a 3D object, requiring heuristic post-merging steps to generate final 3D LN instances. In this study, we formulate 3D LN detection as a tracking task and propose LN-Tracker, a novel LN tracking transformer, that can effectively tackle the challenging LN detection task across major body sections.

Materials/Methods: We collected and curated data from 4 institutions including 1000 patients with 7000+ annotated LNs including various body parts (neck, chest, and abdomen) and different diseases (head & neck, esophageal, lung, and pancreatic cancers). Built upon a transformer-based detector (DETR), LN-Tracker decouples transformer decoder's query into the track and detection groups, where the track query autoregressively follows previously tracked LN instances along the z-axis of a CT scan. We design a new transformer decoder with masked attention and inter-slice similarity loss to align track query's content to the context of current slice, and meanwhile preserving detection query's high accuracy in current slice. We evaluate the LN detection performance on each LN dataset (70% training, 10% validation, and 20% testing), and report the average sensitivity (AS) at 1, 2, 4, 8 FPs per patient and the average precision (AP).

Results: The quantitative evaluation on four LN datasets is summarized in Table 1. As shown, our LN-Tracker surpasses all comparing methods across four datasets, outperforming the closest competitor, LN-DETR (the-state-of-the-art 2.5D LN detector), by 2.7% in mean AS (62.7% vs. 60.0%) and 2.4% in mean AP (52.7% vs. 50.3%). Notably, on the Pan-LN dataset that have thinner z-axis spacing, LN-Tracker achieves significant improvements over the strong 3D detector nnDetection, with markedly margin of 5.1% on AS.

Conclusion: We developed a new whole-body 3D LN tracking transformer, LN-Tracker, which significantly enhances lymph node detection performance with strong generalizability across different body parts and diseases. This model has the potential to play an important role in clinical LN assessment.