Public Chest CT Dataset Supports Explainable AI in Diagnostic Radiology

Clinicians need to verify AI findings before using them in diagnostic radiology workflows. This need is especially important when hundreds of two-dimensional slices must be reviewed to locate a reported abnormality. To help address this challenge, researchers have developed a public dataset that links radiology report descriptions to precise locations in chest CT scans.

The dataset, ReXGroundingCT, was developed by an international team led by biomedical informatics researchers at Harvard Medical School. It connects descriptive radiological language, such as a small lung nodule in a specific lobe, to exact locations in chest CT images. AI models trained with the dataset could produce descriptive reports while linking each finding to specific pixels within a three-dimensional scan.

The approach is designed to support grounding, which is the process of connecting natural-language findings to image locations. For CT scans, this requires pairing accurate descriptions with visual markers that show clinicians where to look across many image slices. The work was published in NEJM AI and addresses the challenge of making radiology AI outputs easier to confirm.

The dataset was created through manual annotation of more than 16,000 anomalies across 3,142 scans. Medical annotators, medical students, and radiologists contributed to the work over one year. Each annotation connected report language to a location in a CT scan and was double-checked by a board-certified radiologist.

The researchers made ReXGroundingCT available to the scientific community ahead of an official challenge at the International Conference on Medical Image Computing and Computer Assisted Intervention, which concluded on September 27. The challenge drew 53 teams from 15 countries. These teams tested different approaches for building models that can link natural language to CT scan locations.

The dataset may also support radiology education and patient-facing tools. The researchers said similar work with annotated X-rays has been used for radiology training. They also hope the CT dataset can help create tools that show patients where findings appear on their own scans and explain associated terminology in plain language.

“My biggest hope is that this trains a generation of models that don't just describe a finding but actually show where it is. A model shouldn't only say, "There's a nodule." It should put a marker on the exact voxel it's talking about, so every statement comes with its evidence,” said Pranav Rajpurkar, associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School.

Related Links
Harvard Medical School