3D densely connected CNN with multi-scale receptive fields and hybrid loss for brain tumor segmentation

Brain tumors, especially gliomas, are among the most common and aggressive types of tumors in the brain. Accurate segmentation of subcortical brain structures is crucial for studying these tumors, monitoring their progression, and evaluating treatment outcomes. However, manual segmentation of magnetic resonance imaging (MRI) data is labor-intensive and limits the use of precise measurements in clinical practice. This paper presents a 3D convolutional neural network (CNN) model to automatically segment brain tumors from MRI scans. A key challenge arises from the significant variations in tumor location, structure, and shape across different patients. Our model addresses this by extracting multi-scale contextual features using two receptive field scales to accurately classify each voxel. A novel hierarchical architecture is proposed to separately delineate the non-enhancing tumor core, peritumoral edema, and enhancing tumor regions, taking full advantage of the tumor structure. Furthermore, densely connected blocks are incorporated to improve feature propagation and boost performance. To mitigate the class imbalance problem, a hybrid loss function combining cross-entropy and Dice loss is employed. The proposed method is evaluated on the MICCAI BraTS (MICCAI BraTS 2017 (Brain Tumor Segmentation Challenge at the Medical Image Computing and Computer Assisted Intervention Conference 2017) is a publicly available dataset of multimodal MRI scans for brain tumor segmentation) 2017 dataset, achieving Dice similarity coefficients of 0.95, 0.89, and 0.89 for the whole tumor, tumor core, and enhancing tumor regions, respectively. These results are on par with the current state-of-the-art while our 3D approach offers improved computational efficiency. The code for this article is available on my GitHub ( https://github.com/fzadib/Brain-Tumor-Segmentation ).