AI Developed to Detect Bone Metastases That Are Easy to Miss on CT... Lesion Detection Precision Reaches 80%

[Sportschosun Jang Jong-ho] Accurately determining whether cancer has spread to the bones is an important step in staging the disease and deciding on a treatment plan.

However, bone metastases are among the lesions that are easy to miss on CT scans. In particular, bone metastases that do not show clear changes on CT are difficult to detect, and delayed diagnosis can lead to serious complications such as pathological fractures or spinal cord compression.

Against this backdrop, a research team led by professors Kim Dong-hyun, Kim Hyo-jin, Seo Ji-woon and Chae Ji-won from the Department of Radiology at Seoul Metropolitan Government Seoul National University Boramae Medical Center, operated by Seoul National University Hospital, developed an artificial intelligence model that learned not only bone metastases clearly visible on CT, but also lesions that are not obvious on CT yet are confirmed as bone metastases on MRI or PET/CT, and found that detection performance improved.

Previous AI studies on bone metastasis detection often used data from a single institution or defined bone metastases based on CT images alone. The problem is that bone metastases that are difficult to detect on CT can be omitted from the training data even when they are real lesions. To address this, the team used MRI and PET/CT as reference imaging as well, including not only lesions clearly visible on CT but also bone metastases that are hard to confirm on CT in the training data.

The study analyzed 502 chest and abdominal CT scans from 332 patients at four medical institutions in South Korea, covering a total of 4,999 bone metastasis lesions. Based on nnU-Net, a 3D image segmentation algorithm, the team developed Model 1, which learned only bone metastases clearly visible on CT, and Model 2, which also learned hard-to-detect bone metastases confirmed through MRI and PET/CT. External validation was then performed using data from a separate medical institution.

As a result, Model 2, which was trained on hard-to-detect bone metastases as well, achieved a lesion-level recall of 41.8%, significantly higher than Model 1's 33.9% (P<0.001). Even when only bone metastases clearly visible on CT were analyzed separately, Model 2's recall was 53.6%, compared with 44.7% for Model 1. Training on more complete ground-truth data, including hard-to-detect lesions, led to better overall bone metastasis detection performance.

The comparison with human readers also showed meaningful performance from Model 2. Its lesion detection precision was 80.1%, higher than 66.6% for the radiology resident group and 66.5% for the musculoskeletal radiology specialist group. Recall was 41.8% for Model 2, 39.4% for the residents, and 43.8% for the musculoskeletal radiology specialists, with no statistically significant difference.

The study showed that improving medical AI performance depends not only on the algorithm, but also on how accurately and comprehensively the training data are built. Rather than defining bone metastases based on CT alone, the researchers used MRI and PET/CT together to include hard-to-detect lesions in the training data, which improved real-world detection performance.

The findings are also significant because they suggest the AI could serve as a safety net, helping radiologists identify bone metastases that might otherwise be missed during interpretation. The AI model analyzed one CT scan in an average of less than 49 seconds, and in the future it could be linked with clinical image viewers to mark suspicious bone metastasis lesions and use segmentation results to quantitatively assess bone metastasis volume and changes before and after treatment.

Professor Kim Dong-hyun, the study's corresponding author, said, "The important point of this study is that we did not train the model only on bone metastases clearly visible on CT images. Instead, we used MRI and PET/CT to build accurate ground-truth data that also included bone metastases difficult to detect on CT." He added, "It is meaningful that an AI model trained on such high-quality data showed bone metastasis detection performance similar to that of musculoskeletal radiology specialists."

He continued, "If the AI model is applied in real CT reading environments in the future, it could serve as an auxiliary safety net by detecting bone metastases that radiologists might miss." He added, "Beyond simply finding lesions, it could be developed further to automatically measure bone metastasis volume and quantitatively evaluate tumor burden and treatment response before and after therapy."

The results of the study were recently published in the international journal Radiology: Artificial Intelligence.

Jang Jong-ho, bellho@sportschosun.com

This article has been translated by GripLabs Mingo AI.