T17: Predicting lesion size measurements


Objective:
Develop a model to extract and predict standardized lesion size measurements from free-text radiology reports. Each report describes one or more lesions and includes a defined prediction objective (e.g., longest diameter, short axis), which is provided as a prefix in the report text.

Patient Population:
Patients with radiology reports from three real-world diagnostic cohorts including:

  • Pulmonary nodules
  • RECIST target lesions
  • Pancreatic ductal adenocarcinomas (PDAC)

Imaging Data:
Not applicable. The task is based solely on textual data — radiology reports written in Dutch.

Test Data:
Each report is provided in a JSON file along with the expected prediction objective. Participants must extract a numerical lesion size in millimeters (mm) according to the provided objective.

Reference Standard:
Lesion size annotations were manually created by trained investigators or radiologists. Ground truth sizes vary by cohort:

  • Pulmonary nodules: 1–30 mm
  • RECIST lesions: 4–166 mm
  • PDAC lesions: 6–130 mm

Evaluation Metrics:
Performance is measured using Robust Symmetric Mean Absolute Percentage Error (RSMAPE) with a 4 mm tolerance, to account for clinical relevance and measurement variability.

Relation to Existing Challenges:

  • Task 17 in the UNICORN challenge is a combination of the following tasks from the DRAGON challenge:

    • T22 (Pancreatic ductal adenocarcinoma size measurement)
    • T23 (Pulmonary nodule size measurement)
    • T24 (RECIST lesion size measurement)
  • Unlike DRAGON, each UNICORN test case includes a prediction objective as a prefix to guide the model, supporting a few-shot learning setup conducted entirely on-platform.

Additional Resources: