Hierarchical anatomical learning improves detection of heart failure signs on chest X-rays
M Kim, J H MoonAbstract
Background
Chest X-ray (CXR) is widely used for initial evaluation and monitoring of heart failure (HF), allowing identification of key signs such as cardiomegaly, pulmonary edema, and pleural effusion. Recent deep learning studies on CXR-based HF assessment have focused on detecting individual signs. However, HF signs on CXR span multiple anatomical regions — the heart (cardiomegaly), lung parenchyma (pulmonary edema), and pleura (pleural effusion) — and frequently co-occur. Conventional approaches that identify each sign independently do not leverage these anatomical relationships, potentially limiting detection performance.
Purpose
To evaluate whether incorporating anatomical context through a hierarchical structure (organ-region-finding) into deep learning improves the detection of HF-related signs on CXR.
Methods
This retrospective study included a total of 240,473 frontal CXR images from 63,458 patients, collected from intensive care units (ICUs) and related clinical settings at an academic medical center. Training and evaluation sets were separated at the patient level, with the evaluation set comprising an independent cohort of 289 patients (3,403 images). A three-level anatomical hierarchy was constructed from annotations derived from radiology reports: organs (heart, lung, pleura) to sub-regions to findings. Three primary HF signs — cardiomegaly, pulmonary edema, and pleural effusion — were designated as primary endpoints. A hierarchical model incorporating anatomical containment relationships between organs, regions, and findings was compared with a model without such hierarchical learning. Performance was assessed using AUC, and model comparison was performed using the DeLong test.
Results
The macro AUC for the three HF signs improved from 0.829 (95% CI 0.820–0.838) with the non-hierarchical model to 0.862 (95% CI 0.854–0.870) with the hierarchical model (+0.033). Consistent improvements were observed across all three signs: cardiomegaly (0.840 to 0.865), pulmonary edema (0.788 to 0.840), and pleural effusion (0.859 to 0.881). The largest improvement was observed for pulmonary edema (+0.052). The hierarchical model showed significantly higher AUC for all three signs (all p<0.001, DeLong test).
Conclusion
Hierarchical anatomical learning improved detection performance across all major CXR signs of heart failure. As HF signs span multiple anatomical regions, incorporating anatomical context complemented the detection of individual findings. These results support the utility of anatomically-informed deep learning for automated heart failure screening on CXR, particularly in high-acuity clinical settings where early detection is essential.