Show abstract - PDF - Reviews

Detection of anomalies before they are included in the downstream diagnosis/prognosis models is an important criterion for maintaining the medical AI imaging model performance across internal and external datasets. However, the core challenges are: (i) given the infinite variations of possible anomaly, curation of training data is in-feasible; (ii) making assumptions about the types of anomalies are often hypothetical. We propose an unsupervised anomaly detection model using a cascade variational autoencoder coupled with a zero-shot learning (ZSL) network that maps the latent vectors to semantic attribute space. We present the performance of the proposed model on two different use cases – skin images and chest radiographs and also compare against the same class of state-of-the art generative OOD detection models.
Hide abstract