Generalizable DNN model for brain tumor sub-structure segmentation from low-resolution 2D multimodal MR Images

Segmenting subregions within gliomas are critical for effective treatment planning of brain tumors. However, traditional methods of analyzing these regions using multiple MRI modalities are time-consuming, tedious, and subjective. To address these challenges, automatic segmentation models have been developed but are often built with complex 3D architecture using 3D MRI data. Also, brain tumor substructure segmentation is a highly class-imbalanced problem. To overcome these limitations, we propose two models that work on low-resolution 2D MRI data, widely used in resource-constrained countries. One model employs training a 2D U-NeT model using proposed hard sampling approach, demonstrating its effectiveness in segmenting gliomas, especially in datasets with extreme class imbalance. Another model incorporates pointwise and depthwise convolutions in each convolutional layer, enabling efficient information processing and feature learning. By ensembling the prediction maps of these models, we further improve overall segmentation performance. Our models were evaluated on the BraTS2018 dataset, achieving dice scores of 0.78 for Enhancing Tumor (ET), 0.82 for Tumor Core (TC), and 0.87 for Whole Tumor (WT). On a tertiary care hospital dataset, dice scores of 0.68 (ET), 0.75 (TC), and 0.84 (WT) were obtained, demonstrating their robustness and proximity to state-of-the-art methods. In summary, the proposed models offer efficient and reliable segmentation of glioma subregions. Their high dice scores, and computational efficiency, make them valuable tools for treatment planning and advancements in brain tumor segmentation.