Multi-class Semantic Segmentation: A Deep Supervised Approach For Brain Tumor Detection Using 3D U-Net++ Architecture

In recent years, the advancement of vision technologies has paved the way for the development of AI applications that can aid in the detection and classification of human brain tumors. Multi-class semantic segmentation is one of the critical applications of such technologies, which has become increasingly important in the field of medical imaging. The detection and accurate segmentation of tumors are crucial for effective diagnosis, treatment planning, and disease monitoring. Multimodal images, such as MRI and CT, provide comprehensive data on the internal structures of the specific part of human body such as brain, spine, lungs, and chest and have become the primary source for medical imaging. This paper primarily focused on the development of an efficient and accurate algorithm for multi-class semantic segmentation in medical imaging for improving the accuracy of tumor detection and diagnosis. The main motive behind this work has been to build a learning approach which would automatically interpret and segment in order to detect, classify, and segment the multiple affected regions of a human brain tumor specifically necrosis, edema, and enhancing tumor. In particular, this paper has explored one specific deep learning algorithm, i.e., 3D U-net++, an extended and improved version of traditional U-Net which is capable of learning deep-hierarchical contextual features, for biomedical image segmentation specifically from brain MRI (BRaTS2020). U-Net++ is incorporated with a deep supervision technique in order to amplify the output from each neuron layer and finally to produce a refined segmented image of the brain tumor. The major reason for choosing this particular neural framework (3D U-Net++) over the popular ones like VGGNet, ResNet, etc. primarily because of the redesigned skip connections which enable highly flexible feature fusion. The working mechanism of the proposed neural architecture has been depicted graphically followed by statistical representations.