Improving Transfer Learning Performance for Abnormality Detection in Brain MRI Images Using Feature Optimization Techniques

Accurate and early diagnosis of brain tumors plays a vital role in determining appropriate treatment and increasing survival rates. Manual diagnosis using MRI images is often prone to errors and can be a tedious and time-consuming procedure. As a result, automated computer-aided systems that utilize artificial intelligence and machine learning techniques are widely used. However, accurately diagnosing the type of brain tumor remains a significant and complex challenge for today’s medical imaging research community. In this paper, we present an enhanced transfer learning approach that aims to differentiate brain tumor types using MRI scans while achieving state-of-the-art performance. In the proposed approach, the pre-trained VGG-19 network with fixed weights is used to map MRI scans to a high-level numerical representation. The extracted features are then passed to a support vector machine algorithm to classify brain tumors into one of three major categories: glioma, meningioma, and pituitary. To further improve the performance of the model, the effectiveness of utilizing feature optimization techniques based on the radial basis similarity function and L1-penalty logistic regression was investigated. For evaluation, the benchmark Figshare dataset containing 3064 MRI images was used after a series of processing operations. The proposed system achieved a high overall accuracy of 98.53%, outperforming other analogues.