Evaluation of the effectiveness of deep learning in the differential diagnosis of vertebral body lesions detected by magnetic resonance imaging

Introduction and Objective: Spinal diseases have become increasingly common with the extension of human life expectancy. Metastases are the most frequently observed malignancies in the vertebral body, while hemangiomas are the most common benign tumors. Furthermore, compression fractures and spondylodiscitis are also frequently encountered. Although MRI is the gold standard for diagnosing these conditions, the similar imaging characteristics of the lesions can pose challenges in the differential diagnosis. In this study, we aims to evaluate the effectiveness of deep learning models in the differential diagnosis of these conditions using MRI. Materials and Methods: Between January 2019 and March 2024, images of patients who underwent thoracic and lumbar spinal MRI using a 1.5 T MR device at Recep Tayyip Erdoğan University Medical Faculty Training and Research Hospital were scanned. A total of 392 vertebral body lesions from 235 patients were included in the study. Suitable image data were obtained from sagittal plane T1 and T2-weighted sequences and recorded. All images were standardized to a uniform size. Two separate datasets were created from the standardized images for T1 and T2 weighted images. Pathology groups were defined using bounding boxes to identify metastasis, acute compression fractures, hemangiomas, atypical hemangiomas, and spondylodiscitis. The images in the created dataset were divided into 80% for training and 20% for validation. Additionally, images from 54 patients were used for external testing purposes. Following the image preprocessing stages, detection and classification were performed using the YOLOv8 deep learning model. Findings: According to the results of the test set, the mAP(B) value for T1 and T2 datasets were 0.82 and 0.86, respectively, and the mAP(M) values were 0.83 and 0.85. The precision(B) values were 0.85 and 0.86, while the precision(M) values were 0.81 and 0.82. The recall(B) values were 0.82 and 0.84, and the recall(M) values were found to be 0.84 and 0.82. The F1 scores were 0.82 and 0.83, respectively. The accuracy rates were 0.84 for the T1 dataset and 0.85 for the T2 dataset. Conclusion: Deep learning models demonstrate high performance in the differential diagnosis of vertebral body lesions. These results indicate that deep learning architectures can be valuable tools in diagnostic processes by performing imaging analyses. Deep learning approaches provide strong support for clinical applications in the detection and differential diagnosis of vertebral body lesions, thereby contributing to the improvement of patient care. Keywords: Vertebral body lesions, magnetic resonance imaging, deep learning, detection, classification, differentional diagnosis