Self-Learning and Evolving X-Ray Imaging Equipment to Be Released

KAIST Develops Technology to Enhance Lung Disease Diagnosis Using Artificial Intelligence

Comparison of attention visualization results of vision transformer models trained with the DISTL method versus previous methods.

[Asia Economy Reporter Kim Bong-su] X-ray imaging equipment that can improve its diagnostic ability by itself using artificial intelligence (AI) is expected to be released.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 27th that a research team led by Professor Yeo Jong-cheol of the Kim Jae-cheol AI Graduate School has developed a self-evolving artificial intelligence technology that can autonomously enhance the automatic reading ability of lung diseases using chest X-ray images such as tuberculosis, pneumothorax, and COVID-19.

Most of the currently used medical AI techniques are based on supervised learning. To train AI models, a large amount of labeling by experts is essential. Continuously obtaining large-scale data labeled by experts in actual clinical settings has been costly and time-consuming, becoming an obstacle to the advancement of medical AI.

The research team developed a self-supervised and self-training AI algorithm (DISTL) utilizing knowledge distillation techniques that describe self-learning and teacher-student knowledge transfer methods, similar to how radiologists learn image interpretation in hospital settings. They demonstrated that by training an initial model with only a small amount of labeled data, the model can improve its performance over time using only the accumulated unlabeled data.

In the actual medical imaging field, the difficulty of acquiring refined labels interpreted by experts frequently occurs regardless of imaging modality or task. This shortage of imaging specialists is common in regions with high incidences of infectious diseases such as tuberculosis, including low-income and developing countries. The AI algorithm developed by the research team is expected to greatly assist in improving diagnostic accuracy by enabling AI models to self-evolve in these regions.

Professor Yeo explained, "To improve performance with supervised learning, continuous acquisition of expert labels is required, and unsupervised learning methods have low performance. We overcame these issues," adding, "This is meaningful in that it surpasses supervised learning performance while reducing the cost and effort of generating labels for AI training by imaging specialists."

The research results were published on the 4th in the international academic journal Nature Communications.

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