SWSSL: Sliding Window-Based Self-Supervised Learning for Anomaly Detection in High-Resolution Images.

Area of Science:

• Medical imaging analysis
• Computer vision
• Machine learning

Background:

• Anomaly detection (AD) is crucial for identifying abnormal instances, but high-resolution medical images pose significant challenges due to exponentially increasing learning difficulty.
• Current methods often compromise diagnostic detail by resizing images, which is unsuitable for clinical practice.

Purpose of the Study:

• To develop an effective anomaly detection method for high-resolution medical images without compromising diagnostic detail.
• To address the challenges of training neural networks at the patch level, such as inconsistent image structure and higher variance.

Main Methods:

• Proposed a patch-level training and inference strategy using a sliding window algorithm to process high-resolution images.
• Focused the network's objective on learning augmentation-invariant features and investigated medical imaging-specific augmentations.
• Introduced a novel module to leverage information from adjacent patches for improved detection performance.

Main Results:

• Achieved significant improvements in anomaly detection accuracy on breast tomosynthesis and chest X-ray datasets.
• Demonstrated an 8.03% and 5.66% increase in Area Under the Curve (AUC) for image-level classification over existing leading techniques.
• Validated the effectiveness of learning augmentation-invariant features and incorporating adjacent patch information.

Conclusions:

• The patch-level approach effectively enables anomaly detection in high-resolution medical images, preserving critical diagnostic details.
• The proposed method, by learning augmentation-invariant features and utilizing contextual patch information, significantly enhances detection performance.
• This technique offers a promising solution for improving diagnostic accuracy in medical imaging analysis.