Unsupervised Deep Anomaly Detection in Chest Radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate anomaly detection in chest radiographs is crucial for timely disease diagnosis.
• Unsupervised methods reduce the need for extensive, expert-annotated datasets.

Purpose of the Study:

• To propose and evaluate an unsupervised anomaly detection method using a deep neural network (DNN).
• To train the DNN model exclusively with normal chest radiograph images.
• To assess the system's performance on a large dataset of chest radiographs.

Main Methods:

• Utilized the auto-encoding generative adversarial network (α-GAN) framework, combining GAN and variational autoencoder principles.
• Trained the model on a dataset of 29,684 frontal chest radiographs, predominantly using normal images.
• Evaluated the system's ability to detect and visualize various thoracic anomalies.

Main Results:

• The anomaly detection system successfully visualized diverse lesions such as lung masses and cardiomegaly.
• Achieved an overall Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.752.
• Demonstrated strong performance with AUROCs of 0.838 for 'Opacity' and 0.704 for 'No Opacity/Not Normal' labels.

Conclusions:

• The proposed DNN-based unsupervised anomaly detection method effectively identifies various diseases and anomalies in chest radiographs.
• Training solely on normal images enables the system to detect deviations indicative of pathology.
• This approach holds promise for improving the efficiency and accuracy of radiological assessments.