ADEEPLEARNINGAPPROACHFORBRAINTUMOR CLASSIFICATION AND SEGMENTATION

Abstract

Brain tumor diagnosis is a critical task that requires accurate and fast identification of the type and location of the tumor. In this paper, we propose using deep-learning models, noise-filtering, and data-augmentation techniques to achieve high performance in brain tu mor classification and segmentation. The brain MRI images have noise due to patient motion, magnetic field inhomogeneities and hardware imperfections. The noise can af fect the accuracy of models trained on those images. We evaluate various noise-filtering techniques and found that non-local means filtering performs the best. Given the limited size of a medical dataset, we perform transfer learning. We assess the impact of various data-augmentation techniques, such as affine and elastic transformations and find the best combinations that lead to the highest performance. Finally, we evaluate several machine and deep-learning models on the “Brain MRI” dataset and observe that VGG16 provides the highest classification accuracy of 98.6%. We further perform segmentation on glioma tumor images from the TCGA-LGG dataset. Weevaluate transformer-based networks (e.g. SegFormer) and CNN-based networks (e.g., U-Net with various encoder backbones). Of these, UNet with ResNet50 provides the high est Dice coefficient of 0.96 and IoU of 0.90. These findings provide valuable insights into the effectiveness of different deep-learning models for brain MRI classification and segmentation. We also underscore the importance of transfer learning, noise-filtering and data-augmentation strategies to extract the full potential of deep-learning techniques. Our research can improve automated brain MRI analysis for clinical diagnosis and treatment planning.