Tumor-Specific Classification of Canine Cutaneous and Subcutaneous Tumors Using Heat-Diffusion Imaging and Artificial Intelligence

Introduction: Cutaneous and subcutaneous tumors are frequently encountered in canine patients, posing diagnostic challenges in general veterinary practice. Heat-Diffusion Imaging (HDI) is an active dynamic thermography technique that evaluates tissue thermal changes following controlled stimulation. Integration of HDI with artificial intelligence (AI) provides a non-invasive approach for tumor classification. Objective: Our objective was to develop and internally validate HDI-based, tumor-specific AI classifiers for differentiating canine mast cell tumors (MCTs) and lipomas from other cutaneous and subcutaneous masses. Methods: This study included 669 dogs with 1010 cutaneous and subcutaneous masses. Dynamic thermal data were recorded from each mass using an HDI system. This was followed by cytological or histopathological diagnosis. Thermal features were extracted from the recordings and used to develop, train, and internally validate a primary AI-based classifier that distinguishes benign from malignant lesions. A diagnostically confirmed subgroup was used to train and internally validate tumor-specific classifiers for lipomas and mast cell tumors. The diagnostic performance of all classifiers was independently evaluated. Results: The primary classifier cohort comprised 952 masses. Of these, 760 masses were assigned to the training cohort and 192 masses to the test cohort. The primary classifier achieved a sensitivity of 92% (95% CI 83.6–96.3) in the test set of 192 masses, and a prevalence-adjusted negative predictive value (NPV) of 98.1% at an assumed malignancy prevalence of 15%. Among the 55 masses assessed by the lipoma classifier, sensitivity was 86.4% (95% CI 73.3–93.6) and specificity 100% (95% CI 74.1–100); all 38 masses flagged as lipoma were lipomas. Among the 81 masses assessed by the MCT classifier, sensitivity was 31.8% (95% CI 20.0–46.6) and specificity 94.6% (95% CI 82.3–98.5); all 16 masses flagged as MCT were malignant, and 14/16 were diagnosed as MCTs. Conclusions: These findings demonstrate that combining HDI-derived thermal features with AI enables promising, non-invasive differentiation of common canine cutaneous tumor types. The two-tiered diagnostic approach improves upon the primary classifier by providing tumor-specific classification. This strategy may be particularly valuable in general practice settings, where access to immediate cytological or histopathological evaluation can be limited.