HYPERPARAMETER ALTERATION IN COMPACT 3D U-NETS FOR BRAIN TUMOR SEGMENTATION

An accurate segmentation of brain tumors is crucial for the treatment of tumor patients, especially for surgical resection. The manual segmentation is very time-consuming and lacks reproducibility, depending on the individual skills of the neuroradiologists. In the recent decade, convolutional neural networks have been deployed to perform this task automatically. However, they are usually computationally expensive and require a long training time. Besides these, there is ambiguity on which hyperparameter’s alteration could enhance the results of tumor segmentation. Therefore, this study aims to investigate the effect of those least explored hyperparameters on the performance of three-dimensional U-Nets in segmenting brain tumors from magnetic resonance images. Seven variants of a residual 3D U-Net were constructed, each with a single altered hyperparameter. All the variants were trained on 400 multimodal magnetic resonance images and tested on another 100 images. Variant 5 was identified as the best-performing 3D U-Net model with the following hyperparameters: 1.0[Formula: see text]mm 3 /voxel resolution, four encoding blocks, excluding one bottleneck block, dropout rates of 0.1, 0.1, 0.2, 0.5, and 0.5 for each stage, and number of convolutional filters of 16, 32, 64, 196, and 256 for each stage. It used the Dice categorical cross-entropy loss function with a Dice: categorical cross-entropy ratio of 1:1. It produced an average Dice score, intersection over union, sensitivity, and specificity of 0.88, 0.82, 0.92, and 0.97, respectively. The proposed model in this study demonstrated comparable performance to previous works while having the advantage of lower complexity with fewer epochs and shorter training time.