search

cnulab / MediCLIP

Official implementation of "MediCLIP: Adapting CLIP for Few-shot Medical Image Anomaly Detection (MICCAI 2024 Early Accept)"
MIT License
28 stars 0 forks source link

about the support set #7

Closed Devil-Ideal closed 4 weeks ago

Devil-Ideal commented 1 month ago

Hi! Thank you for your excellent work. I'm not very familiar with few-shot anomaly detection. I was wondering if using only a few normal images and synthesizing abnormal ones instead of using true abnormal samples directly, is intend to ensure diversity.

cnulab commented 1 month ago

Hello, this is to simulate the real-world scenario where it is difficult to collect abnormal samples from certain areas. The experiment was conducted under the most strictest conditions, where abnormal samples were not available. If real anomalous samples are accessible, the synthesized anomalies can be combined with real anomalies to further improve performance.

Devil-Ideal commented 4 weeks ago

Hello, this is to simulate the real-world scenario where it is difficult to collect abnormal samples from certain areas. The experiment was conducted under the most strictest conditions, where abnormal samples were not available. If real anomalous samples are accessible, the synthesized anomalies can be combined with real anomalies to further improve performance.

thank you for reply