Korean Hospital Team Develops AI Model to Improve Bone Metastasis Detection Using PET/CT and Other Information

A research team at Seoul Metropolitan Government Seoul National University Boramae Medical Center has recently developed an artificial intelligence model capable of identifying bone metastases that are difficult to detect on CT images. By incorporating MRI and PET/CT information during the training process, the model enables lesions that are inconspicuous on CT and prone to being missed to be included in the training data, thereby enhancing detection capability.

Bone metastasis is an important basis for determining cancer staging and formulating treatment plans. If lesions are not clearly visible on CT images, bone metastasis may be detected late and may lead to complications such as pathological fractures and spinal cord compression. Previous related AI studies have largely relied on lesions clearly visible on CT to build training data, and the research team believes that this approach may miss lesions that actually exist but are difficult to identify on CT.

This study included 502 chest and abdominal CT images from 332 patients across four medical institutions in South Korea, with a total of 4,999 bone metastasis lesions annotated. The team developed two AI models based on the three-dimensional image segmentation algorithm "nnU-Net": Model 1 was trained using only bone metastases clearly identifiable on CT; Model 2 was trained by additionally incorporating bone metastases identified via MRI and PET/CT that were difficult to detect on CT. Subsequently, the researchers performed external validation using data from another medical institution.

The results showed that Model 2 achieved a lesion-level recall rate of 41.8%, higher than Model 1's 33.9%, with the difference being statistically significant (P<0.001). Even when analyzing only bone metastases relatively clear on CT, Model 2 outperformed Model 1, with detection rates of 53.6% and 44.7%, respectively.

In comparison with manual reading results, Model 2 achieved a lesion detection accuracy of 80.1%, higher than the 66.6% of the radiology resident group and the 66.5% of the musculoskeletal radiology expert group. However, in terms of lesion detection recall, Model 2 achieved 41.8%, the resident group 39.4%, and the musculoskeletal radiology physician group 43.8%, with no statistically significant differences observed among the groups.

The research team stated that the results demonstrate that the performance of medical AI depends not only on the algorithm but also on the accuracy and completeness of the training data. Supplementing lesions that are difficult to visualize on CT with MRI and PET/CT helps build a more complete training dataset. The AI model takes an average of approximately 49 seconds to complete the analysis of one CT scan. The team anticipates that in the future, it can be integrated into clinical image viewing systems to flag suspicious lesions and further quantify bone metastasis volume and changes before and after treatment through image segmentation results.

Professor Kim Dong-hyun stated that the key to this study lies in using MRI and PET/CT to build a dataset that includes lesions difficult to detect on CT, rather than merely learning from lesions clearly visible on CT. He believes that the AI model trained with high-quality data has demonstrated bone metastasis detection capability comparable to that of musculoskeletal radiologists; if applied in actual CT reading environments, it can serve as an auxiliary safety net to help identify bone metastases that may be missed during the reading process.

The research findings were published in the international journal Radiology: Artificial Intelligence, Volume 8, Issue 3, 2026.