Enhancing Spinal Metastasis Detection and Feature Evaluation on Computed Tomography Scans Using Deep-Learning Systems

Abstract

Spinal metastases can result in pathological fractures, which reduce survival time and quality of life. Physician experience significantly influences the detection of spinal metastases and the evaluation of associated features. This study aims to develop a deep-learning system (DLS) for automatic detection of spinal metastasis and feature evaluation using computed tomography and to determine the impact of the DLS on physician performance in the detection and assessment of spinal metastasis. DLS assistance in a multireader, multicase test study results in higher sensitivity and specificity in spinal metastasis detection and feature evaluation (all p < 0.001). Additionally, resident physicians show a more significant improvement in sensitivity and specificity compared with attending or chief physicians in spinal metastasis detection and most feature evaluation (p < 0.01). In a cohort test study, resident oncologists assisted by the DLS achieve significantly higher sensitivity and specificity compared with those without assistance (all p < 0.01), except for the sensitivity of vertebral body collapse evaluation (p > 0.01). DLS assistance may improve physicians’ performance in the detection and evaluation of spinal metastases, particularly that of resident oncologists.