MMCL-Net: Spinal disease diagnosis in global mode using progressive multi-task joint learning

• Published in: Neurocomputing (Amsterdam) Vol. 399; pp. 307 - 316
• Main Authors: Hong, Yanfei, Wei, Benzheng, Han, Zhongyi, Li, Xiang, Zheng, Yuanjie, Li, Shuo
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 25.07.2020
• Subjects: Densely aggregation; Global optimization; Level-set; Medical image; Multi-structure; Progressive multi-task; 99-00; Level-set; Global optimization; 00-01; Medical image; Progressive multi-task; Multi-structure; Densely aggregation
• DOI: 10.1016/j.neucom.2020.01.112
• ISSN: 0925-2312

•Multiple structures of the spine are directly interdependent and influential, and the multi-tasks under a deep convolutional neural network framework can also influence each other.•Densely dilate ResNet module combat the challenge of size discrepancy and extract significant radiological features.•Integrate the variational level-set function can enrich the local morphological features of the deep learning model.•Applied potential features practical can assist other tasks. Simultaneous detection, segmentation, and classification of multiple spinal structures on MRI is crucial for the early and pathogenesis-based diagnosis of multiple spine diseases in the clinical setting. It is more assistance for radiologists reflections on the disease based on the pathogenesis when the lesion area and its adjacent structures are detected. Obviously, the multiple structures of the spine are directly interdependent and influential, and the multi-tasks under a deep convolutional neural network framework can also influence each other. Multi-task joint optimization in the spinal global mode is a direct outlet to seek the dynamic balance of the above potential correlation. In this paper, we propose a novel end-to-end Multi-task Multi-structure Correlation Learning Network (MMCL-Net) for the detection, segmentation, and classification (normal, slight, marked, and severe) of three types of spine structure: disc, vertebra, and neural foramen simultaneously. And the model is locally optimized to achieve a more stable dynamic equilibrium state. Extensive experiments on T1/T2-weighted MR scans from 200 subjects demonstrate that MMCL-Net achieves high performance with mAP of 0.9187, the classification accuracy of 90.67%, and dice coefficient of 90.60%. The experimental results show that the performance of our method is comparable to that of the state-of-the-art methods.