T13: Classifying pulmonary nodule presence¶
Objective:
Develop a model to determine whether a pulmonary nodule is mentioned in a radiology report. The task involves binary classification (presence vs. absence) based on unstructured Dutch-language report text.
Patient Population:
1,000 unique patients from two Dutch hospitals:
- Radboudumc
- Jeroen Bosch Ziekenhuis
Each patient has one corresponding radiology report.
Imaging Data:
Not applicable. The task is based solely on textual data — radiology reports written in Dutch.
Test Data:
200 anonymized radiology reports reflecting real-world class distributions (i.e., class imbalance). Participants must output a binary prediction (nodule present or absent) for each report.
Reference Standard:
- Binary labels (true/false) manually assigned by a medical student
- The reports were annotated by a medical student under the supervision of a radiologist with 26 years of experience, ensuring quality control despite potential labeling errors.
Evaluation Metrics:
Model performance will be evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC).
Relation to Existing Challenges:¶
- Task 13 is derived from Task002 (Pulmonary Nodule Presence) of the DRAGON Challenge.
- Unlike DRAGON, which supported local training with large labeled datasets, this task supports few-shot training directly on-platform.