Regarding medical datasets

[lixinying0727]

Hello, I would like to know how AdaCLIP and other baseline models perform zero-shot anomaly detection on medical datasets that lack normal samples. How does it distinguish between normal and abnormal data？I hope to receive the author's response, thank you very much!

[caoyunkang]

Currently, these methods can directly perform anomaly detection in arbitrary data, whether from industrial areas or medical fields. It is important to note that the zero-shot performance of these methods is highly dependent on the diversity of the auxiliary annotated datasets and the learning mechanisms employed. Therefore, if similar patterns are included in the training data, we can expect to achieve better performance on the targeted categories.