Anomaly Detection for Medical Images Using Self-Supervised and Translation-Consistent Features.

Area of Science:

• Medical imaging analysis
• Deep learning for anomaly detection
• Computer vision in healthcare

Background:

• Deep learning methods for medical anomaly detection struggle with limited labeled anomalous data, especially for rare diseases.
• Normal medical images are more readily available and easier to collect than anomalous ones.
• Existing methods often require extensive lesion annotations, posing a significant challenge.

Purpose of the Study:

• To propose a novel anomaly detection framework, SALAD (self-supervised and translation-consistent features for anomaly detection).
• To address the challenge of limited labeled anomalous medical images by leveraging abundant normal data.
• To develop a reconstruction-based method that learns normal data manifolds through cross-space translation.

Main Methods:

• SALAD employs an encode-and-reconstruct translation between image and latent spaces to learn normal data manifolds.
• It utilizes two constraints: structure similarity loss and center constraint loss, to ensure translation-consistent and representative feature learning.
• A self-supervised learning module is integrated to enhance anomaly detection accuracy by exploiting information from normal data.

Main Results:

• The proposed framework demonstrates effectiveness in anomaly detection on optical coherence tomography (OCT) and chest X-ray datasets.
• SALAD successfully constructs an anomaly score based on learned self-supervised and translation-consistent features.
• Experimental results validate the approach's capability to distinguish anomalous from healthy medical images.

Conclusions:

• The SALAD framework offers a promising solution for anomaly detection in medical imaging, particularly when labeled anomalous data is scarce.
• Leveraging self-supervised and translation-consistent features enables robust learning from normal data.
• The method shows significant potential for improving diagnostic accuracy in various medical imaging applications.