Enhancing Brain Tumor Classification Through Optimal Kernel Selection With GL 1 $$ {GL}_1 $$ -Regularization

Abstract

Brain tumors, known for their rapid and aggressive growth, are among the most serious and life-threatening diseases worldwide. This makes the development of automated detection methods essential for saving lives. Deep transfer learning has become a highly effective approach for automating brain tumor classification and medical imaging, offering promising solutions on a global scale. However, leveraging a pretrained model typically involves special adaptations. Existing adaptation methods involve freezing or fine-tuning specific layers without considering the contribution level of individual kernels. This work aims to extend the concept of layer-level contributions to the kernel level by employing an adaptive optimization model. Indeed, this paper presents a novel optimization model that incorporates group lasso regularization to control which kernels are frozen and which are fine-tuned. The proposed model selects optimal source features that contribute to the target task. Additionally, the proposed optimization model is solved utilizing proximal gradient descent. The method was evaluated on a three-class brain tumor classification task, distinguishing between glioma, meningioma, and pituitary tumors, using a medical MRI dataset. Several experiments confirm the efficacy of our model in identifying both frozen and fine-tuned kernels, thereby improving data classification. Subsequently, the results obtained are compared with those of state-of-the-art transfer learning methods for comprehensive comparison.