Early diagnosis and classification of brain tumors using deep learning approaches on magnetic resonance images

Brain tumors are among the most lethal types of tumors that can affect individuals of all ages. If left undiagnosed, they can become life-threatening and significantly shorten a person's lifespan. Early and accurate diagnosis is essential for effective treatment, and computer-aided detection systems play a crucial role in assisting medical professionals in identifying and classifying brain tumors. Detecting brain tumors from Magnetic Resonance Imaging (MRI) scans using traditional methods is a challenging and error-prone process that requires the expertise of medical doctors. Incorrect or incomplete diagnoses can lead to severe consequences that endanger human life. To overcome these challenges, many researchers are working on artificial intelligence-assisted autonomous disease detection. In this context, this thesis aims to develop effective and efficient deep learning models for the rapid and reliable diagnosis of brain tumors from MRI scans and to evaluate the performance of these models. In the initial phase of this study, a dataset was created for the thesis research based on the decision of the Fırat University Non-Interventional Ethics Committee, the Scientific Research Permissions Cooperation Protocol with the Elazığ Fethi Sekin City Hospital Directorate, and the commission decision of the Elazığ Provincial Health Directorate. Experiments were conducted, and deep learning models were developed for the study's methodology. However, it was observed that the performance outcomes were low due to the insufficient number of Magnetic Resonance Imaging (MRI) scans in the dataset. To address this limitation, publicly available datasets were explored, and the study proceeded with three different datasets obtained from widely used platforms in the literature, which provide valuable inputs for researchers. Separate experimental processes were carried out and presented under three different sections. In this context, brain image classification was first performed using two different Convolutional Neural Network (CNN) architectures. One of these approaches was VGG-16, a widely used Deep Neural Network model known for its successful classification performance. The model trained with this technology achieved a final accuracy of 80%, classifying images as either "tumor present" or "no tumor". As another deep learning approach, a new and efficient CNN-based model was developed from scratch to classify brain tumor images with high accuracy. This classification included four distinct categories: Glioma, Meningioma, Pituitary, and No tumor. Experiments were conducted during the training, validation, and testing phases using similar parameters, and the results obtained with the optimal parameter tuning were thoroughly compared with other studies in the literature, highlighting the differences. Additionally, an ablation study was performed to demonstrate the robustness and generalizability of the proposed approach, with the results presented in tabular format. The newly proposed Convolutional Neural Network-based model achieved a high classification accuracy of 99.76% in the accuracy evaluation metric and also obtained high values in other evaluation metrics. These results indicate that the proposed model can be used with high accuracy and reliability in brain tumor detection and may provide valuable insights for other research areas. Another and perhaps the most original aspect of this study was the investigation of the impact of transformer-based models on brain tumor classification performance. Transformer-based deep learning models are commonly preferred in many hybrid architectures due to the advantages provided by their flexible structures. However, in the field of image classification, these models tend to exhibit low performance when used with datasets containing a limited number of images. Although some studies have focused on this issue, the effect of distillation techniques on the classification performance of transformer-based models has not been thoroughly explored. Within this context, the aim of this study is to examine the effect of distillation techniques on the classification performance of transformer-based deep learning models used with datasets of limited size. In our study, transformer-based models that do not employ distillation techniques—ViTx32 and ViTx16—were used alongside distillation-based models—DeiT and BeiT. A four-class dataset comprising brain Magnetic Resonance (MR) images was selected for the training and testing processes. As a result of the experiments, it was observed that the distillation-based models DeiT and BeiT outperformed the ViTx16 model, which does not utilize distillation, by 2.2% and 1%, respectively. Additionally, it was found that the use of distillation techniques increased classification performance by approximately 4% in detecting non-tumorous individuals. Furthermore, the training time of each model was analyzed in detail. The results obtained from our study indicate that applying distillation techniques significantly enhances the classification performance of transformer-based deep learning models, especially when working with limited data. Based on the findings of this study, the use of distillation-enhanced transformer-based models is recommended, particularly in the development of flexible models in the medical domain where data availability is often limited.