T7: Detecting lung nodules in thoracic CT¶
Objective:
Develop a model to detect pulmonary nodules in thoracic CT scans from clinical routine and lung cancer screening settings. Participants must output candidate detections: 3D nodule centre coordinates for each predicted nodule and probability score of the candidate being a nodule so [X, Y, Z, P].
Test Data:
Patient Population:
83 patients from Radboudumc, aged 18 and older, who underwent a thorax CT as part of clinical care and were incidentally found to have pulmonary nodules.
Imaging Data:
- Includes both contrast-enhanced and non-contrast scans
- CT volumes in .mha format (512 × 512), intensities already converted to Hounsfield Units (HU)
- No additional preprocessing applied
- Scans obtained from a variety of scanners: GE, Philips, Siemens, and Toshiba
Reference Standard:
Annotations created by a panel of five experienced thoracic radiologists using in-house annotation software. Each pulmonary nodule is annotated with 3D coordinates (center location).
Evaluation Metrics:
- True Positive (TP): A detected candidate within a distance R of a nodule center is considered a TP, where R = nodule diameter / 2
- False Positive (FP): A candidate that does not match any nodule
- Irrelevant Detections: Candidates matching irrelevant findings are ignored in evaluation
Performance is assessed using a Free Receiver Operating Characteristic (FROC) analysis and the final score is the CPM, and is defined as the average sensitivity at 7 predefined false positive rates: 1/8, 1/4, 1/2, 1, 2, 4, and 8 FPs per scan. If none of the candidates result in a true positive, a score of 0 for the CPM will be calculated.
Relation to Existing Challenges:¶
- Task 7 is adapted from the LUNA16 challenge.
- Unlike LUNA16, which focused on local training using a large labeled dataset, this task emphasizes few-shot learning with all training conducted on the platform, and the validation and test sets differ between the challenges.