Hidden-RAD2 consists of two main subtasks. Participants can choose to participate in one or both tasks.
Background: Radiologists interpret chest X-ray images through a structured process, progressing from initial observations to final impressions and confirmation of the supporting evidence. This task evaluates whether AI systems can predict a structured representation of this radiological reasoning.
Input:
A chest X-ray image from the MIMIC-CXR dataset
(Optional) external medical knowledge resources
Output: Radiologist-validated structured reasoning annotations consisting of initial impressions (A1), anatomical locations (A2), thoracic levels or lung zones (A3), final impressions (A4), and an ABCDE-based confirmation checklist (A5).
For further details, see the Task 1 Definition.
Background: While Large Language Models (LLMs) generate fluent text, they can produce "hallucinations" that are disconnected from facts. In the medical field, such errors can be critical. This task is designed to evaluate an AI's ability to self-verify the reliability of its generated text and correct errors. This "critical self-review" capability is essential for ensuring the safety and transparency of AI systems.
Input:
A link to the original radiology report in MIMIC-CXR
A link to the corresponding chest X-ray image in MIMIC-CXR
An AI-generated explanation that may contain seeded errors
The report and image are not distributed by the organizers and must be accessed through an authorized PhysioNet account. Participants may use the report, the image, or both as supporting evidence.
Output:
Error Detection: Identify the location and type of errors within the explanation (e.g., flawed causality, factual inconsistency).
Error Correction: Correct the identified errors with accurate information.
Confidence Score: Provide a confidence score (from 0 to 1) for the overall correctness of the explanation.
For further details, see the Task 2 Definition.