T2: Classifying lung nodule malignancy in CT¶
Objective:
Develop a model that predicts a malignancy risk score for individual lung nodules identified in low-dose chest CT scans. The risk score should indicate the likelihood of a nodule being malignant (non-binarized output).
Test Data:
Patient Population:
Participants from the Danish Lung Cancer Screening Trial (DLCST), recruited between 2004 and 2010. Inclusion criteria:
- Age 50–70 years
- Minimum 20 pack-years of smoking history
Imaging Data:
Low-dose CT thorax-abdomen scans provided as nodule blocks (MHA) sized (128 x 128 x 64), with intensities converted to Hounsfield Units (HU). No additional preprocessing is provided.
- Scans acquired using a 16-slice Philips Mx 8000
- Acquisition: supine position, full inspiration
- Protocol: 120 kV, 40 mA
Reference Standard:
- Malignant nodules: Confirmed via histopathology. For each lung cancer case, the first malignant nodule occurrence was annotated; one malignant nodule per participant was selected randomly.
- Benign nodules: Selected from participants without lung cancer and with at least 2 years of negative imaging follow-up.
- All nodule locations were annotated by two experienced chest radiologists.
Evaluation Metrics:
Area Under the Receiver Operating Characteristic Curve (AUC) is used to evaluate model performance in distinguishing malignant from benign nodules.
Relation to Existing Challenges:¶
- Task 2 is adapted from the LUNA challenge.
- Unlike LUNA25, 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.