Canine radiography and echocardiography sensor fusion for deep learning-based detection and classification of myxomatous mitral valve disease.

BACKGROUND

Early recognition of myxomatous mitral valve disease (MMVD) is crucial for extending the lifespan and improving the quality of life in dogs. Timely diagnosis allows clinicians to initiate medical therapy before the onset of congestive heart failure. However, early-stage MMVD is challenging to identify on radiographs alone, and while echocardiography is highly informative, its interpretation remains highly operator-dependent. This paper describes the development and validation of GGNet, an automated multimodal deep learning framework that synthesizes echocardiographic and thoracic radiographic data for accurate MMVD staging.

METHODS

In this retrospective, single-center study, diagnostic images collected between June 2014 and January 2024 were evaluated. The final cohort comprised 902 dogs (n=902), distributed across stages: normal (n=232), stage B1 (n=243), stage B2 (n=190), and stage C (n=237). Imaging modalities included two-dimensional echocardiograms [right parasternal short-axis at left atrium (LA)/aortic root (Ao) level] and thoracic radiographs [right lateral (RL) and ventrodorsal (VD) projections]. We evaluated GGNet, a dual-input (Echo-RL) architecture employing a high-level feature fusion strategy and a selective partial unfreezing fine-tuning protocol [L2-L4, fully connected (FC) layers], on this dataset. Model performance was evaluated using a five-fold cross-validation approach.

RESULTS

The final GGNet configuration achieved an overall accuracy of 80.3%±2.8%. Crucially, for the clinical task of distinguishing stage B2, the threshold for initiating pimobendan therapy, the model demonstrated a sensitivity of 82.1%±7.5% and a positive predictive value (PPV) of 91.1%±4.0%, effectively minimizing both false negatives and false positives.

CONCLUSIONS

The proposed GGNet framework provides a reliable diagnostic aid for MMVD staging, reducing subjective variability in image interpretation and offering practical decision support for veterinary cardiology. However, its clinical efficacy inherently relies on standardized, high-quality image acquisition by the operator. Furthermore, as a single-center study utilizing internal cross-validation, future external validation on diverse, multi-center datasets is required before broader clinical deployment.