Task 2 Details

Overview

Task 2 evaluates whether an AI system can verify and safely correct a causal explanation generated for a chest X-ray case.

A causal explanation describes how radiological findings and supporting evidence lead to a final diagnostic impression. A hallucination is a claim that is unsupported by the available evidence, clinically incorrect, or internally inconsistent.

Dataset Construction

Each case begins with a clinically reviewed gold causal explanation derived from a structured chest X-ray interpretation and the corresponding medical report.

One or a small number of controlled errors may be inserted into selected claims. The dataset also includes valid explanations without inserted errors.

The errors may involve:

• factual consistency;

• anatomical location or laterality;

• severity;

• negation;

• causal reasoning;

• missing supporting evidence; or

• internal contradictions.

Available Data

All authorized participants receive:

• a causal exploration section;

• the original MIMIC-CXR radiology report; and

• the corresponding MIMIC-CXR chest X-ray image.

Participants may decide which evidence to use during system development. They do not need to register for an evidence category in advance.

Evidence-Use Categories

R — Report-only

The submitted system uses the causal exploration section and the original radiology report. It does not use the image during inference.

I — Image-only

The submitted system uses the causal exploration section and the chest X-ray image. It does not use the original report during inference.

RI — Report-and-Image

The submitted system uses the causal exploration section, original report, and chest X-ray image.

These categories identify the evidence used by a submitted run. They are not separate subtasks.