Explainable AI to identify radiographic features of pulmonary edema.

Abstract

Background: Pulmonary edema is a leading cause for requiring hospitalization in patients with congestive heart failure. Assessing the severity of this condition with radiological imaging becomes paramount in determining the optimal course of patient care.

Purpose: This study aimed to develop a deep learning methodology for the identification of radiographic features associated with pulmonary edema.

Materials and methods: This retrospective study used a dataset from the Medical Information Mart for Intensive Care database comprising 1000 chest radiograph images from 741 patients with suspected pulmonary edema. The images were annotated by an experienced radiologist, who labeled radiographic manifestations of cephalization, Kerley lines, pleural effusion, bat wings, and infiltrate features of edema. The proposed methodology involves 2 consecutive stages: lung segmentation and edema feature localization. The segmentation stage is implemented using an ensemble of 3 networks. In the subsequent localization stage, we evaluated 8 object detection networks, assessing their performance with average precision (AP) and mean AP.

Results: Effusion, infiltrate, and bat wing features were best detected by the Side-Aware Boundary Localization (SABL) network with corresponding APs of 0.599, 0.395, and 0.926, respectively. Furthermore, SABL achieved the highest overall mean AP of 0.568. The Cascade Region Proposal Network network attained the highest AP of 0.417 for Kerley lines and the Probabilistic Anchor Assignment network achieved the highest AP of 0.533 for cephalization.

Conclusion: The proposed methodology, with the application of SABL, Cascade Region Proposal Network, and Probabilistic Anchor Assignment detection networks, is accurate and efficient in localizing and identifying pulmonary edema features and is therefore a promising diagnostic candidate for interpretable severity assessment of pulmonary edema.