Improve the performance of CT-based pneumonia classification via source data reweighting.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pneumonia diagnosis relies heavily on computed tomography (CT) imaging.
• Deep learning models offer potential for automated pneumonia detection in CT scans.
• Acquiring large annotated CT datasets for training is challenging due to privacy and cost.

Purpose of the Study:

• To develop a method for improving pneumonia detection in CT scans when labeled data is scarce.
• To leverage data from a source domain to enhance performance in a target domain with limited annotations.
• To address the limitations of existing deep learning methods requiring extensive labeled data.

Main Methods:

• A three-level optimization strategy was employed to utilize source domain CT data.
• The method automatically identifies and downweights low-quality or discrepant source data examples.
• Optimization involved minimizing the validation loss of a target model trained on reweighted source data.

Main Results:

• The proposed method achieved an F1 score of 91.8% for general pneumonia detection.
• An F1 score of 92.4% was obtained for detecting other types of pneumonia.
• Performance significantly surpassed state-of-the-art baseline methods on the target dataset.

Conclusions:

• The developed method effectively mitigates the challenge of limited labeled CT scans for pneumonia detection.
• Leveraging and optimizing source domain data improves the accuracy and efficiency of deep learning models in medical imaging.
• This approach offers a promising solution for enhancing radiological diagnoses in data-scarce scenarios.