DEEP LEARNING APPROCHES FOR TUMOR DETECTION USING MRI DATA

Neurological diseases are relatively severe in the field of health informatics, often associated with life-threatening symptoms and costly treatments. Among these, brain tumors stand out as a well-known concern, showing a noteworthy increase in the affected patient’s number over the past decade. MRI imaging is the primary method for tumor detection, and recent advancements in Computer-Aided Diagnosis (CAD) using deep learning have improved diagnostic accuracy. However, existing models have drawbacks, such as inadequate dataset sizes, which hinder early-stage tumor detection, and the limited number of extracted features from input images. To address the challenges, an Enhanced Convolutional Neural Network (ECNN) model has been proposed. The proposed ECNN is trained on the MRI images taken from the BR35H benchmark dataset with extensive data augmentation techniques to improve generalization. The ECNN model achieved a high accuracy of 99.3% in classifying tumor images. Once tumor-positive images were identified, further analysis was performed using a Vision Transformer (ViT) model, trained on a different subset of the BR35H dataset. The ViT model achieved an accuracy of 97% in localizing tumor regions, showing its effectiveness in precise tumor segmentation. This hybrid technique, which addresses important issues in automated brain tumor diagnosis, improves both detection accuracy and interpretability by using CNN-based classification followed by Transformer-based localization.