Advancements in brain tumor classification: a survey of transfer learning techniques

Abstract

This survey article presents a critical review of the state-of-the-art transfer learning (TL) methodologies applied in the field of brain tumor classification, with a special emphasis on their various contributions and associated performance metrics. We will discuss various pre-processing approaches, the underlying fine-tuning strategies, whether used purely or in an end-to-end training manner, and multi-modal applications. The current study specifically highlights the application of VGG16 and residual network (ResNet) methods for feature extraction, demonstrating that leveraging highorder features in magnetic resonance imaging (MRI) images can enhance accuracy while reducing training. We further analyze fine-tuning methods in relation to their role in optimizing model layers for small, domain-specific datasets, finding them particularly effective in enhancing performance on the small brain tumor dataset. It will look into end-to-end training, which means fine-tuning models that have already been trained on large datasets to make them better. It will also present multimodal TL as a way to use both MRI and computed tomography (CT) scan data to get better classification results. Comparing different pre-trained models can provide a better understanding of the strengths and weaknesses associated with the particular brain tumor classification task. This review aims to analyze the advancements in TL for medical image analysis and explore potential avenues for future research and development in this crucial field of medical diagnostics.