Automated detection of central chest lymph nodes for guided bronchoscopy

Physicians now commonly use guided bronchoscopy systems to help diagnose lung cancer in early stage patients. To use such a system, the physician draws on a patient’s high-resolution chest computed tomography (CT) scan to preplan a bronchoscopic procedure for a peripheral nodule observed on the CT. Later, in the operating room, the physician uses the guidance system, driven by the procedure plan, to navigate the bronchoscope to the nodule and perform the desired tissue biopsy. Central-chest lymph node staging, however, is also required to ascertain a patient’s disease stage. Unfortunately, current practice tends to rely on manually identifying only 2 or 3 enlarged lymph nodes (short axis ≥ 10 mm) on a thick-slice CT scan (slice spacing ≥ 3 mm). Yet, to have a truly accurate and effective assessment of disease stage, the staging procedure should satisfy two criteria: 1) consider a more comprehensive set of nodes distributed across the chest; 2) include nodes smaller than what had heretofore been considered in clinical practice. Because manual identification of lymph nodes is known to be challenging, researchers have suggested automatic deep learning methods for CT-based lymph extraction. Unfortunately, these methods have only focused on defining enlarged lymph nodes observed in thick-slice CT scans, as typically done by radiologists. We present a three-stage framework that detects lymph nodes across a broader size range; i.e., short axis ≥ 5 mm; and draws on high-resolution chest CT scans used for guided bronchoscopy (slice spacing < 1 mm). The framework includes: 1) pre-processing, to reduce redundant information; 2) model inference, using a trained Swin-UNETR Transformer model; and 3) post-prediction filtering, to eliminate false positives. On a 41-patient training and validation dataset containing 847 nodes having short axis ≥ 5 mm, our model showed an average sensitivity of 76.5% for lymph node detection. In a separate comparison test involving a set of 3D high-resolution chest CT scans, our framework gave detection rates of 80%, 91%, and 71% for all nodes, enlarged nodes and smaller nodes, respectively, versus 52%, 73%, and 38% for a state-of-the-art deep learning method. Furthermore, the proposed framework achieved 85.4% detection rate on a widely used public dataset with 388 enlarged mediastinal lymph nodes from 30 CT scans.