Task 1:
Structured Radiological Reasoning Prediction 

Overview

Hidden-RAD2 Task 1 evaluates whether an AI system can predict the radiologist-validated structured reasoning annotations associated with a chest X-ray image. Rather than generating only a final diagnosis or a free-text report, the system must predict the intermediate impressions, anatomical locations, thoracic levels or lung zones, final impressions, and confirmation evidence represented by A1–A5.

Background

During dataset construction, a radiologist examines a chest X-ray image, referred to as Source 1 (S1), and completes a structured interpretation consisting of five components:

1. A1 — Initial Impressions: All possible impressions considered during the initial image review

2. A2 —  Anatomical Location: The anatomical location of each finding

3. A3 — Thoracic Level: The relevant thoracic spine level or lung zone

4. A4 — Final Impressions: The final diagnostic decisions after reviewing the available evidence

5. A5 — Confirmation Checklist: An ABCDE-based checklist identifying the evidence that supports each final impression

The corresponding medical report, referred to as Source 2 (S2), is used as a reference during annotation. However, some diagnostic evidence or causal relationships may not be explicitly documented in the report. The A5 checklist captures this supporting information and makes the reasoning behind each final impression more explicit.

NTCIR-18 Hidden-RAD Task 1 required participants to generate a causal exploration section from an existing radiology report, with optional use of the corresponding chest X-ray image. The objective was to recover the hidden causal relationships connecting the findings described in the report to its diagnostic impressions.

Hidden-RAD2 extends this task by moving from report-based explanation generation to image-based prediction of radiologist-validated structured reasoning annotations.

Task 1 Definition

Given a chest X-ray image, a participating system must predict the corresponding radiologist-validated A1–A5 structured interpretation.

The system must:

• identify all reasonable initial impressions;

• identify the anatomical locations of the findings;

• determine the relevant thoracic levels or lung zones;

• decide the final impressions, including a normal conclusion when appropriate; and

• provide an ABCDE-based confirmation checklist supporting each final impression.

Input 

Each test case provides:

• • A chest X-ray image from the MIMIC-CXR source dataset

Access to MIMIC-CXR images requires the appropriate PhysioNet credentials and data-use authorization.

Participants may optionally use external medical knowledge resources, pretrained models, or other supporting information, subject to the task rules.

Required Output

For each test image, participants must submit the following structured output:

A1 — Initial Impressions

List all possible impressions considered during the initial image review. These are preliminary candidates and do not necessarily need to appear in the final impressions.

A2 — Anatomical Location

Identify the anatomical location and, when applicable, the laterality of each finding.

A3 — Thoracic Level

Specify the relevant thoracic spine level or lung zone associated with each observation.

A4 — Final Impressions

Provide the final diagnostic impressions supported by the image evidence. A case may be classified as normal even when some preliminary observations were considered in A1.

A5 — Confirmation Checklist

Complete the ABCDE-based checklist for each final impression. The checklist must identify the findings and clinical evidence that confirm or support the diagnostic decision.

Training Data

Each training case contains:

• a chest X-ray image; and

• a radiologist-validated A1–A5 structured interpretation.

The structured annotations capture both the final diagnostic conclusions and the intermediate reasoning used to reach them.

Objective

The primary objective is to predict a transparent and clinically meaningful structured representation of radiological reasoning. A successful system should accurately predict not only the final impressions but also the associated anatomical information, intermediate impressions, and confirmation evidence.