Deep Learning-Based Differentiation of Vertebral Body Lesions on Magnetic Resonance Imaging

Abstract

Objectives: Spinal diseases are commonly encountered health problems with a wide spectrum. In addition to degenerative changes, other common spinal pathologies include metastases and compression fractures. Benign tumors like hemangiomas and infections such as spondylodiscitis are also frequently observed. Although magnetic resonance imaging (MRI) is considered the gold standard in diagnostic imaging, the morphological similarities of lesions can pose significant challenges in differential diagnoses. In recent years, the use of artificial intelligence applications in medical imaging has become increasingly widespread. In this study, we aim to detect and classify vertebral body lesions using the YOLO-v8 (You Only Look Once, version 8) deep learning architecture. Materials and Methods: This study included MRI data from 235 patients with vertebral body lesions. The dataset comprised sagittal T1- and T2-weighted sequences. The diagnostic categories consisted of acute compression fractures, metastases, hemangiomas, atypical hemangiomas, and spondylodiscitis. For automated detection and classification of vertebral lesions, the YOLOv8 deep learning model was employed. Following image standardization and data augmentation, a total of 4179 images were generated. The dataset was randomly split into training (80%) and validation (20%) subsets. Additionally, an independent test set was constructed using MRI images from 54 patients who were not included in the training or validation phases to evaluate the model's performance. Results: In the test, the YOLOv8 model achieved classification accuracies of 0.84 and 0.85 for T1- and T2-weighted MRI sequences, respectively. Among the diagnostic categories, spondylodiscitis had the highest accuracy in the T1 dataset (0.94), while acute compression fractures were most accurately detected in the T2 dataset (0.93). Hemangiomas exhibited the lowest classification accuracy in both modalities (0.73). The F1 scores were calculated as 0.83 for T1-weighted and 0.82 for T2-weighted sequences at optimal confidence thresholds. The model's mean average precision (mAP) 0.5 values were 0.82 for T1 and 0.86 for T2 datasets, indicating high precision in lesion detection. Conclusions: The YOLO-v8 deep learning model we used demonstrates effective performance in distinguishing vertebral body metastases from different groups of benign pathologies.

Keywords: MRI; YOLO-v8; classification; deep learning; detection; metastasis; vertebral body lesions.

Figure 1

Dataset preparation: (a) T1-weighted MRI image of the spine, (b) classification and labeling of groups by the radiologist (red box), (c) image augmentation using PIL, including cropping (0–20% zoom), rotation (−15° to 15°), shearing (±15° horizontally and ±15° vertically), noise addition, and brightness adjustment.

Figure 2

Workflow of the study.

Figure 3

YOLO-v8 architecture.

Figure 4

Comparison of YOLO-v8 and previous versions based on the COCO dataset graph [16].

Figure 5

An example demonstrating the detection and classification of spinal lesions using bounding boxes with the YOLOv8 model. In the figure, each bounding box contains the detected lesion type (e.g., acute compression fracture, atypical hemangioma, and metastasis) along with the model’s confidence score for the prediction.

Figure 6

Graphs showing the loss values and validation metrics related to the performance training analysis of YOLO-v8.

Figure 7

Classification accuracies based on the dataset obtained from T1-weighted images are as follows: 0.89 for acute compression fracture, 0.80 for metastasis, 0.73 for hemangioma, 0.84 for atypical hemangioma, and 0.94 for spondylodiscitis.

Figure 8

Classification accuracies based on the dataset obtained from T2-weighted images are as follows: 0.93 for acute compression fracture, 0.89 for metastasis, 0.73 for hemangioma, 0.79 for atypical hemangioma, and 0.86 for spondylodiscitis.

Figure 9

The Precision–Confidence curve, F1 Score–Confidence curve, Precision–Recall curve, and Recall–Confidence curve for the T1 dataset are shown.

Figure 10

The Precision–Confidence curve, F1 Score–Confidence curve, Precision–Recall curve, and Recall–Confidence curve for the T2 dataset are shown.