Improving Efficiency of Spinal Cord Image Segmentation Using Transfer Learning Inspired Mask Region-Based Augmented Convolutional Neural Network

Abstract

Spinal cord magnetic resonance images (MRIs) consists of 7 levels of cervical vertebrae, 12 levels of thoracic vertebrae, 5 levels of lumbar vertebrae, one level each of sacrum and coccyx components. Segmentation of these components is essential for effective classification and post-processing analysis of spinal cord images. In order to perform this task, separate algorithms are needed for each of the components. Due to which, their segmentation performance is not uniform, which limits their integration capabilities. Moreover, performance for each type of segmentation has scalability issues, which must be improved via augmentation, aggregation, and machine learning for better clinical use. To resolve these issues, this text proposes design of a novel spinal cord image segmentation model using transfer learning inspired mask region-based augmented convolutional neural network (MRACNN). The proposed model utilizes initial weights from the pre-trained COCO mask RCNN model, and modifies them to incorporate spine, torso, and ${\text{L}}_{1}$ to ${\text{L}}_{5}$ spinal cord components. When compared to several state-of-the-art models, it is found that the suggested model has improved region of interest (RoI) extraction and an accuracy of 91% for segmenting these components. Moreover, the proposed model was evaluated on multiple datasets, and a consistent performance was observed. Furthermore, the model was fused with a XRAI-based convolutional neural network which assisted in further improving overall efficiency of segmentation. Fusion of XRAI CNN with MRACNN is capable of achieving segmentation accuracy of 94%, along with better RoI performance when compared with individual models. Delay requirement of the fused model is high, and requires large dataset for training & validation, thus, this text also recommends selective ensembling techniques for redundancy reduction, which assists in improving segmentation speed, while maintaining high segmentation quality.