Abstract Rationale Accurate differentiation between central and peripheral lung nodules is critical for selecting appropriate surgical candidates, as recent randomized trials (JCOG 0802, CALGB 140503) have established sublobar resection as equivalent to lobectomy for small (≤2 cm), peripheral nodules. However, translating these findings into practice requires accurate tumor localization, which currently relies on manual CT interpretation—a process that is time-consuming, subjective, and inconsistent across institutions. We developed and validated an automated ray-casting algorithm for lung wedge segmentation that provides objective, reproducible peripheral nodule classification using physician-defined hilar landmarks. Methods We analyzed 290 annotated nodules from the LIDC-IDRI dataset. Lung and airway segmentation was performed using publicly available deep learning models. Left and right hilum points were manually annotated by a physician. Our ray-casting algorithm employed Fibonacci sphere sampling to generate 500,000 uniformly distributed rays emanating from each hilum. Using 3D Digital Differential Analyzer (DDA) traversal, each ray identified the farthest lung boundary point, defining the outer third wedge as the region from 2/3 of the hilum-to-boundary distance to the lung edge. We compared two classification approaches: (1) simple overlap (any nodule-wedge overlap = peripheral), and (2) volumetric threshold (≥50% nodule volume in wedge = peripheral). Automated classifications were validated against manual annotations. Results The simple overlap method achieved 93.4% concordance with manual annotations (19/290 discrepancies). All 19 discrepant cases were manually classified as central but automatically classified as peripheral, representing false positives for peripheral classification. The 50% volumetric threshold method demonstrated superior performance with 96.9% concordance (9/290 discrepancies). Of the 9 discrepancies, 5 were manual central/automated peripheral and 4 were manual peripheral/automated central, showing more balanced error distribution. The improved specificity of the volumetric method reduced false peripheral classifications by 52.6% compared to simple overlap. Conclusion Our automated ray-casting algorithm for lung wedge segmentation demonstrates excellent concordance with manual nodule classification, with the 50% volumetric threshold achieving 96.9% agreement. This method provides an objective, reproducible approach for peripheral nodule identification that could standardize surgical candidate selection for sublobar resection and reduce inter-observer variability in clinical practice. Future work will focus on automated hilum detection and refinement of chest wall nodule classification. This abstract is funded by: U.S. Department of Veterans Affairs, Biomedical Laboratory Research and Development (VA BLR&D), Grant I01BX004121-05A1
Al-Shakhshir et al. (Fri,) studied this question.