Early identification of abnormal pulmonary infectious diseases using unsupervised anomaly detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Early identification of abnormal pulmonary infectious diseases (APIDs) is crucial for controlling disease spread.
• Deep learning offers a promising approach for automated disease detection.

Purpose of the Study:

• To propose a deep learning-based method for the early identification of APIDs.
• To develop and evaluate a novel unsupervised anomaly detection (UAD) network for detecting deviations in pulmonary infection computed tomography (PICT) image sequences.

Main Methods:

• Established a PICT image sequence dataset including common pulmonary infections and specific abnormal cases (COVID-19, melioidosis pneumonia).
• Trained a UAD network using only common infection sequences and tested its ability to identify deviations in both common and abnormal sequences.
• Developed the local reconstruction autoencoder (LRAE) to enhance detection accuracy by focusing on local image regions.

Main Results:

• The LRAE successfully identified known abnormal cases (COVID-19, melioidosis pneumonia) and detected unknown APIDs.
• Achieved high performance metrics: AUC, F1-score, and accuracy (ACC) for COVID-19 detection were 0.8269, 0.7242, and 0.7801, respectively.
• Demonstrated superior performance compared to existing UAD methods on the PICT dataset, with AUC, F1-score, and ACC for melioidosis pneumonia detection at 0.8716, 0.6415, and 0.8146, respectively.

Conclusions:

• The developed LRAE provides a robust solution for the early identification of both known and emerging APIDs.
• The LRAE demonstrated remarkable performance in detecting abnormal PICT image sequences, highlighting its clinical potential.