Tri-VAE: Triplet Variational Autoencoder for Unsupervised Anomaly Detection in Brain Tumor MRI

The intricate manifestations of pathological brain lesions in imaging data pose challenges for supervised detection methods due to the scarcity of annotated samples. To overcome this difficulty, our focus shifts to unsupervised anomaly detection. In this work, we exclusively train the proposed model using healthy data to identify unseen anomalies during testing. This study entails investigating the triplet-based variational autoencoder to simultaneously learn the distribution of healthy brain data and denoising capabilities. Importantly, we rectify a misconception inherent in prior projection-based approaches which relies on the presumption that healthy regions within images would persist unaltered in the reconstructed output. This inadvertently implied a substantial likeness in latent space representations between lesion and lesion-free images. However, this assumption might not hold true, particularly due to the potential significant impact of lesion area intensities on the projection process notably for autoencoders with single information bottleneck. To overcome this limitation, we disentangled metric learning from latent sampling. This approach ensures that both lesion and lesion-free input images are projected into the same distribution, specifically the lesion-free projection. Moreover, we introduce a semantic-guided gated cross skip module to enhance spatial detail retrieval while suppressing anomalies, leveraging robust healthy brain representation semantics exist in the deeper levels of the decoder. We also discovered that incorporating structure similarity index measure as an extra training objective bolsters the capability of anomaly detection for the proposed model.