Comparison Performance of Deep Learning Models for Brain Tumor Segmentation Based on 2D Convolutional Neural Network

A brain tumor is a lump of aberrant brain cells whose formation can interfere with the brain's normal functioning and adversely affect the patient's health. Brain tumor segmentation is a fundamental step for quantitatively analyzing tumor masses and extracting infected areas of brain tissue. Accurate and reliable segmentation is helpful in clinical diagnosis and treatment planning so that the probability of survival can be improved. The deep learning technique showed a significant performance improvement in brain tumor segmentation. The objectives of this study are to comprehensively investigate and compare the effectiveness of multiple deep learning based on 2-dimensional Convolutional Neural Networks for clustering brain tumors on MRI images. This study implements a simple U-Net model and modifies it with different backbones as encoders, including VGG19, ResNet50, Inception-V3, and InceptionResnetV2. This model was trained and tested on the BraTS 2020 dataset. The tumor area was segmented into three main areas: necrotic, edematous, and enhancing tumors. The predicted image is weighed against the ground truth for validation. The predictive performance of each model was analyzed using quantification metrics. The results showed that the five models tested, including Simple U-Net, VGG19-UNet, Res-UNet, InceptionV3-UNet, and InceptionResnetV2-UNet, achieved 98.8% to 99.3% pixel accuracy. However, when the evaluation was carried out on the three sub-areas, each model showed comparable performance, except for the VGG19-UNet model, which performed significantly lower. It shows that performance can be influenced by many factors, such as the amount of training data, loss function, parameter tuning, and the hyperparameters of each model. The proposed model can be implemented well. Thus, this research needs to be developed further.