Extracting Disease Names from X-rays... Kakao Brain Unveils Labeler Project

Kakao Brain announced on the 19th that it has released a specialized medical imaging report labeler project that extracts specific disease names from chest X-ray reports on 'GitHub'.

Kakao Brain CI [Image provided by Kakao Brain]

The labeler project is a research and development project that extracts specific disease names from unstructured natural language reports, such as those written in bullet point format. It is characterized by the ability to extract a total of 13 disease names with high incidence or importance among diagnosable diseases based on chest X-ray results, including ▲fractures ▲pleural lesions ▲pneumothorax. For example, when a user inputs a report written in natural language, the system analyzes the report and indicates the presence or absence of the 13 disease names.

Kakao Brain initiated the labeler project research to accurately and efficiently extract specific disease names in order to contribute to the improvement of medical diagnostic work. This project has also been used for internal performance verification research of Kakao Brain’s chest X-ray draft report generation technology.

Kakao Brain measured the accuracy of disease name extraction targeting 10 actual diseases, including ▲fractures ▲pneumothorax ▲pulmonary edema. As a result, Kakao Brain’s labeler project recorded a significantly higher accuracy of 90.39% compared to other company models (approximately 76%).

Kakao Brain strives to share AI technology know-how and create new value in line with Kakao and its community’s core value of 'a better world created by technology.' Similar to the open-source project 'Honeybee (tentative name)' released last January, this labeler release is also aimed at revitalizing the AI open-source ecosystem.

Ildu Kim, Co-CEO of Kakao Brain, said, "We plan to release a test set we created ourselves so that many researchers can use Kakao Brain’s labeler project as a test benchmark," adding, "We plan to further improve the performance of the labeler project by utilizing our language model and additionally training it with chest X-ray data."

This content was produced with the assistance of AI translation services.