Artificial Intelligence for Diagnostic Interpretation of Small Animal Thoracic Radiographs: A Review of Methods, Clinical Evidence, and Future Directions

Artificial intelligence is rapidly entering small-animal thoracic radiography, yet methodological choices, evidence quality, and deployment practices remain heterogeneous. This narrative review examines approaches across supervised multi-label classification, segmentation-based measurement systems that compute vertebral heart score (VHS) and cardiothoracic ratio (CTR), transformer models that regress clinically interpretable key points, and upstream quality assurance for positioning, collimation, and exposure. We appraise evidence by clinical task and ground truth, emphasizing echocardiography for cardiac endpoints, thoracocentesis or ultrasound for pleural effusion, and blinded expert adjudication for cardiogenic pulmonary edema. Findings across species show a consistent hierarchy: models perform best on high-contrast targets such as pleural effusion, pneumothorax, and marked cardiomegaly; performance declines for bronchial and interstitial textures and for masses, where label scarcity, heterogeneity, and domain shift are pronounced. Automated VHS and CTR agree with specialists, and geometry-aware transformer overlays improve trust through auditable measurements. Quality assurance at acquisition and repository-level AutoQC mitigate shortcut learning and stabilize downstream performance. In contrast, many studies rely on single-center datasets, image-level labels, and limited external validation, which constrains generalizability. Nevertheless, self-supervised learning on same-modality radiographs and multi-view pretraining using paired projections improve robust representations. The most credible path to impact is a layered pipeline that begins with quality assurance, proceeds through interpretable measurements or localized detections, and culminates in calibrated, clinician-in-the-loop decisions. Priorities include multicenter, patient-level validation with confirmed endpoints, multi-view inputs, rigorous calibration with confidence intervals, and reader-assist trials that measure workflow and outcomes, not accuracy.

کلیدواژه ها:

Small-animal thoracic radiography ، artificial intelligence ، cardiomegaly ، pleural effusion ، multi-view self-supervised learning