Development of an artificial intelligence-based algorithm for the detection of left atrial enlargement from feline thoracic radiographs

Comprehensive Summary

Valente et al. developed a convolutional neural network (heart-CNN) to identify left atrial enlargement (LAE) in cats using routine thoracic radiographs, with echocardiography serving as the reference standard. The retrospective study analyzed 694 radiographs from 270 cats across three institutions, incorporating both right lateral and dorsoventral/ventrodorsal projections and classifying LAE severity using established left atrium/aorta thresholds. When tasked with four-class severity grading, the model performed assymetricly: discrimination was acceptable for cats with no LAE and those with severe enlargement but was notably weaker for separating mild from moderate disease (AUC range: 0.63-0.78). But performance improved when the task was simplified to identifying clinically meaningful disease, with binary classification of moderate-to-severe versus no-to-mild LAE achieving AUCs of 0.83 (right lateral) and 0.81 (DV/VD), and accuracies of 81% and 79% respectively. Overall, the model appears better suited to flagging advanced LAE than fine-grained staging.

Outcomes and Implications

In practice, this model is best viewed as a supplement to radiographic interpretation rather than a stand-alone diagnostic tool. In settings where echocardiography is unavailable or delayed, such as general practice or emergency care, AI-assisted radiograph analysis could help identify cats with moderate-to-severe LAE who would benefit from expedited cardiologic evaluation. At the same time, the model’s limited ability to distinguish intermediate disease severity reinforces that echocardiography remains essential for definitive staging and management decisions. Before clinical adoption, prospective validation and comparison with experienced radiologist interpretation will be important to clarify how such tools should be integrated into veterinary workflows.