Vision transformer-based unsupervised anomaly detection in chest CT using feature-space masked prediction

Reconstruction-based Vision Transformer (ViT) models enable unsupervised anomaly detection in medical images without lesion-level labels, but lesion-related information can remain in latent patch features and be reconstructed, reducing residual anomaly signals. To address this limitation, we propose a ViT-based framework incorporating feature-space masking, in which each target patch feature is masked while retaining its learnable target positional embedding and is predicted from spatially neighboring features using a dedicated mask predictor. Chest CT volumes are reconstructed from these neighbor-conditioned predicted features, reducing direct leakage of target-patch information while preserving anatomical context. We evaluated the proposed method on 1,050 chest CT cases, using only normal cases for training and validation. For case-level anomaly detection, the proposed method achieved an AUROC of 0.851, significantly outperforming 3D-AnoViT (0.694, adjusted P < 0.001), 3D-MAE (0.800, adjusted P = 0.029), and f-AnoGAN (0.684, adjusted P < 0.001). Using a validation-defined threshold, the proposed method showed high disease-specific sensitivities for the evaluated abnormality categories, including lung cancer, emphysema, pneumonia, and pneumothorax. These findings suggest that feature-space masking with neighbor-conditioned prediction improves reconstruction-based unsupervised anomaly detection in a balanced chest CT evaluation cohort.