Deep learning model for pleural effusion detection via active learning and pseudo-labeling: a multisite study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pleural effusion affects 1.5 million annually in the US, necessitating timely diagnosis.
• Current diagnostic methods require timely and accurate detection of pleural effusion.
• Development of clinical-grade algorithms for pleural effusion detection is critical.

Purpose of the Study:

• To develop and validate a deep learning-based Computer Aided Triage (CADt) algorithm for pleural effusion detection.
• To utilize an active learning (AL) framework to enhance the efficiency of CADt algorithm development.
• To address the need for a clinical-grade algorithm for timely pleural effusion diagnosis.

Main Methods:

• A deep learning algorithm was trained on 10,599 chest radiographs from Taiwan (2006-2018).
• An active learning (AL) framework was employed to minimize the requirement for expert annotations.
• External validation was performed on 600 chest radiographs from 22 clinical sites in the US and Taiwan.

Main Results:

• The CADt algorithm achieved high diagnostic performance: sensitivity 0.95, specificity 0.97.
• The area under the receiver operating characteristic curve (AUC) was 0.97, indicating excellent accuracy.
• The algorithm demonstrated robust performance across diverse demographics and clinical settings.

Conclusions:

• A novel, active learning-based CADt algorithm for pleural effusion diagnosis was successfully developed and validated.
• The AL-based CADt system achieved high accuracy while significantly reducing the annotation workload for clinical experts.
• This approach enhances the feasibility of utilizing advanced technology for prompt and accurate medical diagnoses.