A Comparative Analysis of Single-Modality and Multi-Modality Multi-View MR Images for Semantic Segmentation: A Study on Brain Tumor Annotation and Segmentation using 3D Convolutional Network

Introduction: Gliomas are well known for their dismal outcomes with limited treatment options. Precise segmentation is essential to develop automated pipelines for a personalized glioma management approach. By utilising cutting-edge methods like multi-view fusion and 3D convolutional neural networks (CNN), the study seeks to increase the precision of brain tumour segmentation.

Methods: The study used T1 weighted MR images of glioma patients from Aga Khan University Hospital (AKUH) (n=100) and the BraTS 2020 dataset (n=460) for data collection and interpretation. A single-modality deep learning model for brain tumour segmentation was proposed. The multi-view model was refined using local datasets and a single modality of BraTS 2020. A 3D convolutional neural network-based encoder and decoder network was used for feature extraction. The proposed model’s performance was assessed, using Hausdorff distance, sensitivity, specificity, and DICE similarity coefficient (DSC).

Results: The multi-view model design performed reasonably well, achieving DSC values of 0.77, 0.89, and 0.78 for Enhancing Tumor (ET), Whole Tumor (WT), and Tumor Core (TC) respectively on the BraTS 2020 validation dataset, along with sensitivity scores of 0.77, 0.87, and 0.75 for ET, WT, and TC), respectively, specificity of 0.99 for ET, WT, and TC. On ET, WT, and TC respectively, and the Hausdorff distances were 3.87, 5.39, and 6.79, respectively.

Conclusion : The study's findings indicate that the suggested strategy trained on BrATS 2020 dataset showed promise for segmenting brain tumours on a locally collected dataset of glioma MRIs.