Explainable AI to identify pulmonary edema

Customer: Beth Israel Deaconess Medical Center (Boston, United States)

Summary: This study introduces a deep learning methodology for identifying radiographic features associated with pulmonary edema, a common cause of hospitalization in congestive heart failure patients. Using a dataset of 1000 chest radiographs from 741 patients, annotated by a radiologist, the research employs a two-stage process: lung segmentation and edema feature localization. The segmentation utilizes an ensemble of three networks, while the localization stage evaluates eight object detection networks based on average precision (AP) and mean AP (mAP). The Side-Aware Boundary Localization (SABL) network demonstrated superior performance in detecting effusion, infiltrate, and bat wing features, achieving an overall mAP of 0.568. The Cascade Region Proposal Network excelled in identifying Kerley lines, and the Probabilistic Anchor Assignment network was best for cephalization. This approach offers an accurate, efficient, and interpretable tool for assessing pulmonary edema severity, promising enhanced diagnostic capabilities in clinical settings.

Collaborators: Anton Makoveev (Quantori, Prague, Czech Republic), Alexander Proutski (Quantori, Hague, Netherlands), Yuriy Gankin (Quantori, Cambridge, United States)

Project type: Commercial / Research

Media:  Journal paper,  GitHub repo

Figure 1. Schematic representation of the proposed approach. The original chest x-ray is fed into the first segmentation stage where a lung mask is predicted. The chest x-ray is then cropped using the lung mask and is further fed into the localization stage to identify radiological features that can be used for further clinical analysis.