Sora Owada

Unsupervised anomaly detection (UAD) in brain MRI aims to learn the distribution of healthy anatomy from unlabelled scans and flag deviations as potential lesions. This avoids the need for voxel-wise annotations and can, in principle, detect a wide range of pathologies, making it attractive for screening and quality-assurance roles in neuroimaging. However, existing UAD work spans many model families—reconstruction-based autoencoders and VAEs, feature-distance and one-class methods, generative models, and hybrids—while using inconsistent evaluation protocols, datasets, and thresholding rules. This makes it difficult to compare results fairly or understand how different methods actually trade off sensitivity and false positives in realistic settings. In this project, I will build a minimal but rigorous UAD benchmarking setup focused on clarity and reproducibility rather than architectural novelty. Concretely, I will (1) select one healthy MRI dataset and one public lesion dataset, (2) perform basic but standardized preprocessing and patient-level train/validation/test splits with no leakage, (3) implement two simple UAD baselines—a reconstruction-based autoencoder/VAE and a Pseudo-Healthy Generative Models trained only on healthy data—and (4) reports voxel-level ROC–AUC and Dice, plus scan-level ROC–AUC. Thresholds will be chosen using a held-out healthy validation set to control false positives, rather than tuned directly on lesion labels.