Unsupervised lesion detection via image restoration with a normative prior

Area of Science:

• Medical imaging
• Artificial intelligence
• Machine learning

Background:

• Unsupervised lesion detection is crucial for identifying abnormalities without labeled data.
• Deep learning advances enable more accurate estimation of healthy anatomical distributions.
• Current methods using prior-projection often yield high false positive rates.

Purpose of the Study:

• To develop a novel approach for unsupervised lesion detection.
• To address the high false positive rates in existing methods.
• To improve the accuracy of detecting lesions in medical images.

Main Methods:

• Framing unsupervised lesion detection as an image restoration problem.
• Proposing a probabilistic model with a network-based prior for normative distribution.
• Utilizing Maximum A Posteriori (MAP) estimation for pixel-wise lesion detection.

Main Results:

• The proposed method significantly reduces false positives compared to prior-projection techniques.
• Achieved a substantial performance improvement of +0.13 AUC for both glioma and stroke detection.
• MAP-based image restoration proved effective in enhancing detection accuracy.

Conclusions:

• The novel probabilistic model effectively detects lesions pixel-wise while minimizing false positives.
• This approach offers a significant advancement over current state-of-the-art unsupervised lesion detection methods.
• The findings demonstrate the potential of image restoration techniques in medical image analysis.