Dataset

We provide a dataset constructed specifically for the Hidden-RAD2 challenge, derived from the MIMIC-CXR-JPG database on PhysioNet and the IU-Xray dataset.

Data Composition:

• Training Data:

  • Task 1: Includes approximately 2,000 reports (as a goal) paired with gold-standard explanations written by radiologists (from Hidden-Rad challenge in NTCIR-18 and plus new cases).

  • Task 2: Includes approximately 300 AI-generated explanations containing seeded errors, paired with the ground truth for error locations and corrections. The errors cover a range of types and difficulties, from minor inconsistencies to clinically significant mistakes.

• Test Data:

  • Will be released later in the challenge period. It will consist of approximately 500 instances with the ground-truth labels held out.

Annotation Process:

The gold-standard explanations in the dataset are authored by radiologists following a structured guideline to ensure high quality and consistency. This process involves documenting initial findings, anatomical locations, final impressions, and a checklist of causal factors for the diagnosis.

Refer to https://sites.google.com/view/ntcir18-hidden-rad/data for Task 1.