Comparison of Machine Learning Techniques for Segmentation of Spinal Cord Tumors from MR Images

Abstract

Precise detection and segmentation of spinal cord tumors from magnetic resonance imaging (MRI) assume a crucial part in boosting clinical examination applications. To capture intricate tumor details, recent techniques have made use of multi-modal imaging, which incorporates the T1, T1c, T2, and FLAIR modalities. While a considerable lot of these methodologies show promising outcomes on benchmark datasets like Brats2018, hurdles continue to arise because of the inherent intricacy, demanding significant training and testing times. A novel strategy is proposed to improve the adaptability and efficiency of spinal cord tumor segmentation in response to these difficulties. To start, a Cascade Deep Learning model's overfitting and computational demands are reduced by introducing a preprocessing method that focuses only on a small area of the image. In this way, an original Cascade Convolutional Brain Network (C-ConvNet/C-CNN) is introduced, decisively intended to remove both local and global features through unique routes. To additionally further develop segmentation accuracy past existing state-of-the-art models, a pivotal Distance-Wise Attention (DWA) system is presented. The DWA component considers the spatial connection between the tumor location and the encompassing brain structures, upgrading the accuracy of growth depiction. The proposed model's efficacy is demonstrated by extensive testing on the BRATS 2018 dataset. The outcomes feature competitive results, with mean whole tumor, enhanced tumor, and tumor core dice scores of 0.9203, 0.9113, and 0.8726, separately. Quantitative and qualitative assessments are presented and discussed, highlighting the effectiveness of the proposed strategy in spinal cord tumor segmentation from MRI images. This exploration adds to the continuous endeavours in propelling AI methods for clinical image investigation, offering an important stratagem for clinicians and scientists in the field.