Anomaly Detection in Brain Tumor Imaging: Few-Shot Learning with Generative Models and Knowledge Transfer
A. Khodabakhsh, Ali Aghaee, Naghmeh Niroomand, Hossein Arabi, Habib Zaidi · 2024
Anomaly detection plays a vital role in the early identification of brain tumors in MRI scans, and it directly impacts diagnostic accuracy and patient outcomes. Despite its importance, current methods often fall short in handling sparse labeled data and precisely localizing anomalies. In this study, an innovative method that integrates Generative Adversarial Networks (GANs) with few-shot learning and transfer learning techniques, offering a new perspective in handling scarce labeled data in medical image analysis. At the core of the proposed method is a Vision Transformer-based generator, showcasing the dedication to advancing medical diagnostics technology. This generator, paired with a uniquely adapted discriminator benefiting from a pre-trained VGG16 network, enhances the model’s efficiency and accuracy in anomaly detection. The efficacy of the proposed method is demonstrated through its ability to distinguish between normal and patho- logical brain images obtained from a public dataset. In our study, the model achieved anomaly scores of approximately $0.14 \pm 0.18$ for normal images and $0.23 \pm 0.76$ for abnormal images. The enhanced precision achieved in anomaly detection and localization signifies a notable advancement beyond current methodologies. These outcomes pave the way for the creation of increasingly sophisticated and dependable diagnostic instruments, thereby facilitating more precise detection of brain tumors.