Comparative Analysis of Vision Transformers and CNN-based Models for Enhanced Brain Tumor Diagnosis

Patient care quality is directly linked to the accuracy and success of diagnostic radiography, notably in the diagnosis and treatment of brain tumors. Central to both medical treatments used for diagnosis and treatment, segmentation serves as a fundamental precursor. Recent advancements in machine learning aim to improve the effectiveness of diagnosis by transitioning from specialized applications to generalized algorithms. Unsupervised anomaly detection plays a crucial role in identifying possible areas of anomaly, particularly those not observed during instruction. Autoencoders based on convolutional neural networks (CNN) and variational autoencoders have emerged as standard approaches for segmenting anomalies based on reconstruction. However, CNNs' constrained receptive fields impose limitations on their ability to capture global context, thereby hindering semantic understanding when large sections of the image are filled with anomalies. We investigate the possibilities of vision transformers (ViTs) in this research as a viable alternative to CNNs for brain magnetic resonance imaging abnormality segmentation. ViTs depend on self-attention mechanisms to establish relationships between image patches, offering a more comprehensive view of the image. We present a comparative study wherein five transformer-based models are developed to reconstruct coherent and realistic images for anomaly segmentation. Our findings demonstrate that these ViT-based models achieve segmentation effectiveness that is on par with or surpassing cutting-edge CNN-based methods. By harnessing the capabilities of ViTs, we pave the way for improved anomaly detection and segmentation in brain MRI, thereby enhancing diagnostic accuracy and patient care in radiology practice.