Self-supervised Deep Learning Model for Generalized Brain Anomaly Detection

Introduction: Medical imaging is pivotal in diagnosing brain anomalies due to its high resolution and non-invasive nature. Detecting anomalies such as malignancies, demyelination, and atrophy with artificial intelligence can be challenging due to the amount of labeled data (scarcity and cost) required and the complexity of having multiple disease processes occurring concurrently (e.g., atrophy and lesions). In contrast, self-supervised techniques, which require only disorder-free images, offer a promising alternative that may also be more generalizable to more complex patients. This study focuses on enhancing anomaly detection using a combination of U-Net reconstruction and Beta-VAE architecture to model normal brain structures and identify deviations.

Methods: The model was trained on disorder-free images obtained from the Stanford BrainMetShare project, consisting of 156 pre-and post-contrast whole-brain MRI studies. Scans confirmed as disorder-free were used, with datasets divided for training, validation, and testing. An external Alzheimer’s disease (AD) dataset was also used to evaluate generalization capabilities. Anomalies were detected using high-dimensional hypercube thresholding to identify latent space deviations.

Results: Using a sliding “abnormality” threshold, we assess the performance of our model. Our model achieved a sensitivity of 95.95% and a specificity of 98.94% at the 99% threshold range. At the 98% threshold, sensitivity increased to 97.34% with a specificity of 97.94%. The 95% range reached a sensitivity of 98.71% with a specificity of 94.95%. Our model significantly outperforms the sensitivity of 83% from the baseline convolutional neural network in the original study.

Conclusion : The model provides a novel disorder-free reconstruction approach for anomaly detection, effectively generalizes to diverse datasets, and offers a flexible, reliable classification method. This approach supports accurate anomaly identification without requiring annotated data, making it suitable for settings with limited labels. Future work will focus on enhancing generalizability and robustness across more complex disorders.