,
Department of Radiology Techniques, Mosul Medical Technical Institute, Northern Technical University (NTU), Mosul, Iraq, Department of Biomedical Engineering and Health Sciences, Faculty of Electrical Engineering, Universiti Teknologi Malaysia (UTM), Johor, Malaysia
,
Sport Innovation and Technology Centre (SITC), Universiti Teknologi Malaysia, Skudai, Malaysia, Department of Biomedical Engineering and Health Sciences, Faculty of Electrical Engineering, Universiti Teknologi Malaysia (UTM), Johor, Malaysia
,
IJN–UTM Cardiovascular Engineering Centre, Institute of Human Centered Engineering, Universiti Teknologi Malaysia, Johor Bahru, Johor, Malaysia, Department of Biomedical Engineering and Health Sciences, Faculty of Electrical Engineering, Universiti Teknologi Malaysia (UTM), Johor, Malaysia
,
Department of Information Technology Management, Technical College of Management, Northern Technical University (NTU), Mosul, Iraq
,
Department of Radiology, Mosul Medical College, University of Mosul, Iraq
Radiology Unit, Department of Surgery, Nineveh Medical College, University of Nineveh, Iraq
Low back pain can be caused by lumbar disc disorders; however, correct segmentation of the structures of the vertebrae and the disc disorders using MRI is crucial for clinical evaluation. In this case, previous deep-learning segmentation approaches have mainly concentrated on the segmentation of anatomical structures and lacked pathology-aware disc segmentation ability. Therefore, a modified attention U-denseNet model for concurrent lumbar anatomical structure segmentation and pathology-aware disc segmentation is proposed in this work using MRI. The approach uses multiscale feature extraction, a channel-spatial attention mechanism, and semantic context gating to increase the performance of the network in discriminating between healthy and abnormal parts of the discs. In this research, a 13-class segmentation task is formulated using 264 sagittal lumbar MRI images. Training, validation and testing were conducted by separating the data based on the subject-wise partitioning strategy to avoid data leakage. Experimentation revealed that the developed model performed with a macro dice of 0.8440, a macro IoU of 0.7985, a boundary F1-score of 0.8693, and a Hausdorff distance of 10.5245 pixels. The performance of the proposed model was better compared to that of the MANet baseline, which yielded a macro dice of 0.7754 and disc condition segmentation with 0.486 dice. The results obtained reveal that semantic context modelling can improve the pathologic awareness of segmentation and help radiologists to make decisions. The proposed framework provides an effective computer-aided segmentation approach; however, further validation using larger multi-centre datasets and three-dimensional MRI analysis is required.
This is an open access article distributed under the Creative Commons Attribution Non-Commercial License (CC BY-NC) License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The statements, opinions and data contained in the journal are solely those of the individual authors and contributors and not of the publisher and the editor(s). We stay neutral with regard to jurisdictional claims in published maps and institutional affiliations.
