Generalized Anomaly Detection in Medical Imaging with Vision-Language Models

• Author / Creator

  Bao, Jinan

• In the realm of image processing, machine learning models have achieved remarkable progress in tasks such as classiffcation, recognition, and video analysis. However, their reliance on closed-set assumptions limits their performance in real-world scenarios where unseen anomalies frequently occur. This limitation is particularly critical in medical applications, where undetected anomalies can have severe consequences. Addressing this open-set problem through anomaly detection is imperative for developing robust systems capable of adapting to unpredictable real-world scenarios.
  In this context, medical anomaly detection stands as a pivotal challenge due to the inherent unpredictability of pathological conditions. Despite recent advances, existing benchmarks for anomaly detection primarily focus on industrial and natural
  images, neglecting the speciffc requirements of medical domains. This has led to inconsistencies in data utilization and a lack of standardized evaluation protocols, impeding fair comparisons among methods. To bridge this gap, we introduce BMAD, a uniffed benchmark tailored speciffcally for assessing anomaly detection methods on medical images. BMAD comprises six reorganized datasets spanning ffve medical domains, alongside standardized evaluation metrics and a comprehensive codebase
  supporting 15 state-of-the-art algorithms.
  The analysis on BMAD reveals that no single model achieves universal effectiveness across multiple medical domains, emphasizing the need for more generalized approaches. To this end, we propose a novel multimodal framework leveraging the
  Contrastive Language-Image Pre-training (CLIP) model to identify anomalies within medical data. By employing language models to describe and reconstruct image information , our approach achieves state-of-the-art performance across multiple domains, demonstrating improved generalization capabilities.

• Subjects / Keywords

  • Anomaly detection
  • Medical benchmarks
  • Vision-Language Model

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-dhns-bg21

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.