Brain Tumor Segmentation on MRI Images Using 2D ResNeXt

This study proposes using 2D Residual Networks with Exponential Number of Transformations (ResNeXt) architecture for brain tumor segmentation in Magnetic Resonance Imaging (MRI) images. This research aims to improve accuracy and speed up computational efficiency during training. The 2D ResNeXt model has advantages in accuracy and computational time efficiency that can be utilized for segmentation tasks, as it utilizes the advantages of residual structure and cardinality concepts. We trained the 2D ResNeXt model and compared it with other ResNet models using the same MRI dataset. The dataset used in this study is BraTS2020, which has been converted from 3D to 2D. Experimental results show that 2D ResNeXt outperforms in several evaluation metrics, with accuracy values on train data 99.5000% and 99.1065% val data. Loss value 0.0198% on train data and 0.0424% on val data. Dice accuracy 99.3606% train data and 98.9745% val data. Dice Loss 0.0063% data train and 0.0102% data val. Mean IoU 0.8863. While the computation time in the training process is 51 minutes and 8 seconds with 100 epochs. The results of this experiment increase the accuracy by 0.0208% and dice accuracy 0.0424%. At the same time, the computation time increased by 9 minutes and 57 seconds faster. The experiments that have been conducted show that the 2D ResNeXt model can be applied to brain tumor segmentation tasks with excellent performance, and the results obtained are more accurate in predicting images.