Unsupervised anomaly localization in high-resolution breast scans using deep pluralistic image completion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated tumor detection in Digital Breast Tomosynthesis (DBT) faces challenges due to rare tumor occurrences and data variability.
• Existing anomaly localization methods often perform poorly on medical imaging datasets.
• Image completion offers a promising approach, but precise anomaly detection is hindered by multiple valid normal image completions.

Purpose of the Study:

• To develop an effective anomaly localization method for Digital Breast Tomosynthesis (DBT) datasets.
• To address the limitations of existing anomaly detection techniques in medical imaging.
• To improve the precision of tumor detection by exploring pluralistic image completion.

Main Methods:

• Proposed an anomaly detection approach from an image completion perspective.
• Introduced pluralistic image completion by exploring the distribution of possible completions.
• Utilized spatial dropout during inference for diverse completions without additional training.
• Developed the Minimum Completion Distance (MCD) metric for anomaly detection.

Main Results:

• The proposed method effectively addresses the challenge of multiple valid normal completions.
• Spatial dropout during inference generated diverse image completions.
• The Minimum Completion Distance (MCD) metric demonstrated superiority for anomaly localization.
• Achieved at least a 10% AUROC improvement over state-of-the-art methods on the DBT dataset for pixel-level detection.

Conclusions:

• Pluralistic image completion with spatial dropout and the MCD metric offers a robust solution for anomaly localization in DBT.
• The novel approach significantly enhances the accuracy of automated tumor detection in medical imaging.
• This method provides a promising direction for improving diagnostic tools in breast cancer screening.