Improved Classification of Brain-Tumor MRI Images Through Data Augmentation and Filter Application

초록

Early diagnosis and treatment of brain tumors can lead to high cure rates. Specialists typically use magnetic resonance imaging (MRI) images to diagnose tumors, and accurate diagnosis is crucial for initiating early treatment. This study proposes a computer-aided specialized system for brain-tumor classification using MRI images that can assist specialists in diagnosing brain tumors. We used an object classification algorithm based on deep-learning networks and augmented the dataset using the ImageNet augmentation policy of AutoAugment to optimize brain-tumor classification performance. Gaussian filtering was applied to enhance the quality of MRI images by removing white Gaussian noise; the parameter value that exhibited relatively high performance in MRI images was determined by varying the sigma values of the Gaussian filter. To evaluate the performance of the proposed technique, we used EfficientNetV2-L(21 K) based on a convolutional neural network and ViT-L/16(21 K) based on transformers. The proposed technique improved the performance of both models. EfficientNetV2-L (21 K) achieved a performance improvement of 2.87% and an accuracy of 0.9590, while ViT-L/16 (21 K) achieved a performance improvement of 4.46% and an accuracy of 0.9413.

키워드