Tracking temporal progression of benign bone tumors through X-ray based detection and segmentation

Comprehensive Summary

This study presents an AI system that automatically measures and compares benign bone tumor (BBT) size and shape on serial X-rays, addressing the current reliance on subjective, manual longitudinal assessment. FusionX-BBTNET’s strengths lie in optimized object detection and precise feature-enhanced segmentation to accurately outline tumor boundaries from plain X-rays. Tumor localization was performed using an optimized YOLO-based object detection model. When compared to expert-annotated ground truth, the model closely matched expert markings with a mean accuracy exceeding 98% and a Boundary F1 score of 0.9827, enabling reliable measurement of the tumor-bone interphase (zone of transition). Further, with the X-ray scale bar on each X-ray image, a real-world distance per pixel conversion was made, allowing the model to achieve an impressive average deviation of 0.46% from expert manual size measurements. Finally, interpretability and visualization are enhanced through centroid-based alignment and color-coded contour overlays on longitudinal X-ray images. This enables rapid visual assessment of tumor stability or growth. Unlike prior models that focus on single time-point classification and detection, FusionX-BBTNET allows for quantitative, real-world longitudinal tracking of tumor size and shape using standard X-ray images.

Outcomes and Implications

BBTs often indicate that the tumor will not metastasize, yet they can still lead to cortical thinning and increased fracture risk if they enlarge. As such, clinicians often need to make long-term observations of tumor size and shape changes, yet most comparisons remain manual and objective quantification models are still lacking. Limitations include single-center data analysis, dependence on consistent X-ray positioning in 3D space, and presence of scale bars which are not universally present in routine imaging. However, with modifications, FusionX-BBTNET can reduce inter-observer variability and support more objective follow-up decisions for long-term surveillance.