Diagnostic accuracy of artificial intelligence technologies in analyzing magnetic resonance imaging results for detecting degenerative changes in the lumbosacral spine

Degenerative spinal diseases are a major medical problem and a common cause of chronic pain, reduced quality of life, and loss of work capacity. Modern treatment approaches, including conservative methods and surgical interventions, do not always demonstrate sufficient clinical efficacy, highlighting the need for improved diagnostic capabilities. Magnetic resonance imaging is the primary modality for visualizing spinal pathologies, providing a detailed assessment of structures, including intervertebral discs, the spinal cord, and surrounding tissues. However, routine MRI diagnostics face significant methodological challenges, such as the need to perform numerous standardized measurements. The integration of artificial intelligence into radiology offers a solution to these problems: algorithms can automate image processing, speeding up analysis and enhancing diagnostic accuracy. Objective:  To evaluate the diagnostic accuracy of automated detection of degenerative changes in the lumbosacral spine using AI-assisted MRI. Materials and Methods. A retrospective diagnostic study was conducted, comprising 100 examinations. Artificial intelligence services (hereinafter referred to as AI service 1, AI-2, AI-3) integrated into the Unified Radiological Information System (URIS) of the Moscow Unified Medical Information and Analytical System (EMIAS) were used as the index test. A test dataset was prepared by three experts based on MRI scans of the lumbosacral spine (T2 sagittal, T2 axial, 4 mm slice thickness). Inclusion criteria were patients aged 18–90 years without artifacts or comorbid pathologies. The evaluation of the AI services involved analyzing degenerative changes: dorsal intervertebral disc protrusions; measurements (anteroposterior diameter of the dural sac on axial and sagittal images, transverse diameter of the dural sac on axial images); and the cross-sectional area of the dural sac. Conclusion . Calibration testing confirmed the importance of evaluating the effectiveness of AI services in medicine. AI-1 and AI-3 demonstrated high reliability (AUC >0.95), making them suitable for clinical use. AI-2 requires refinement, as its performance metrics were significantly lower than claimed and further declined in external validation, limiting its practical application.