Sayed Md Abu, Rahman G M Mahmudur, Islam Md Sherajul, Islam Md Alimul, Park Jeongwon, Ahmed Hasan, Hossain Akram, Shahrior Rahat
Khulna University of Engineering and Technology, Khulna, 9203, Bangladesh.
Department of Electrical and Biomedical Engineering, University of Nevada, Reno, NV, 89557, USA.
Sci Rep. 2025 Jan 2;15(1):593. doi: 10.1038/s41598-024-84301-7.
Magnetic resonance (MR) images are commonly used to diagnose prolapsed lumbar intervertebral disc (PLID). However, for a computer-aided diagnostic (CAD) system, distinguishing between pathological abnormalities of PLID in MR images is a challenging and intricate task. Here, we propose a comprehensive model for the automatic detection and cropping of regions of interest (ROI) from sagittal MR images using the YOLOv8 framework to solve this challenge. We also propose weighted average ensemble (WAE) classification and segmentation models for the classification and the segmentation, respectively. YOLOv8 has good detection accuracy for both the lumbar region (mAP50 = 99.50%) and the vertebral disc (mAP50 = 99.40%). The use of ROI approaches enhances the accuracy of individual models. Specifically, the classification accuracy of the WAE classification model reaches 97.64%, while the segmentation model achieves a Dice value of 95.72%. This automatic technique would improve the diagnostic process by offering enhanced accuracy and efficiency in the assessment of PLID.
磁共振（MR）图像常用于诊断腰椎间盘突出症（PLID）。然而，对于计算机辅助诊断（CAD）系统而言，在MR图像中区分PLID的病理异常是一项具有挑战性且复杂的任务。在此，我们提出一种综合模型，利用YOLOv8框架从矢状面MR图像中自动检测和裁剪感兴趣区域（ROI），以应对这一挑战。我们还分别提出了加权平均集成（WAE）分类模型和分割模型用于分类和分割。YOLOv8对腰椎区域（mAP50 = 99.50%）和椎间盘（mAP50 = 99.40%）均具有良好的检测精度。使用ROI方法提高了各个模型的准确性。具体而言，WAE分类模型的分类准确率达到97.64%，而分割模型的Dice值达到95.72%。这种自动技术将通过在PLID评估中提供更高的准确性和效率来改善诊断过程。
