Anatomy-aware self-supervised learning for anomaly detection in chest radiographs.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Unsupervised anomaly detection (UAD) in medical imaging is crucial for identifying abnormalities without labeled data.
• Existing UAD models often struggle with subtle anomalies in complex anatomical structures like chest radiographs.
• Self-supervised learning (SSL) offers a promising avenue for UAD by leveraging unlabeled data.

Purpose of the Study:

• To develop and evaluate a novel SSL-based model for unsupervised anomaly detection in chest radiographs.
• To introduce an anatomy-aware augmentation technique (AnatPaste) for pretraining UAD models.
• To demonstrate the effectiveness of incorporating anatomical information into SSL for improved UAD performance.

Main Methods:

• A self-supervised learning (SSL) model was developed for unsupervised anomaly detection (UAD).
• An anatomy-aware pasting (AnatPaste) augmentation tool was utilized, employing a lung segmentation pretext task to generate realistic anomalies in normal chest radiographs.
• The model was pretrained using these generated anomalies and then evaluated on three public chest radiograph datasets.

Main Results:

• The proposed SSL-based UAD model achieved high performance across three datasets, with area under the curve (AUC) values of 92.1%, 78.7%, and 81.9%.
• These AUC values represent the highest performance among existing UAD models evaluated on the same datasets.
• The AnatPaste augmentation method demonstrated its effectiveness in improving the model's ability to detect anomalies.

Conclusions:

• This study presents the first SSL model utilizing anatomical information from segmentation as a pretext task for UAD.
• The integration of anatomical context within SSL significantly enhances the accuracy of unsupervised anomaly detection in chest radiographs.
• The findings suggest a promising direction for developing more robust and accurate AI-driven diagnostic tools in radiology.