CT-Based Radiomic Features Predict Cervical Lymph Node Metastasis in Dogs With Oral Malignancy: A Machine Learning Study Using Leave-One-Patient-Out Cross-Validation.
Authors
Affiliations (10)
Affiliations (10)
- Department of Oncology, Toronto Animal Cancer Centre, Toronto, Ontario, Canada.
- ANI.ML Research, ANI.ML Health Inc., Toronto, Ontario, Canada.
- Department of Small Animal Clinical Sciences, University of Saskatoon, Saskatoon, Saskatchewan, Canada.
- Centre for Advancing Responsible and Ethical Artificial Intelligence (CARE-AI), University of Guelph, Guelph, Ontario, Canada.
- Department of Clinical Studies, Ontario Veterinary College, University of Guelph, Guelph, Ontario, Canada.
- Department of Engineering, Anderson University, Anderson, Indiana, USA.
- Radiogenomics Laboratory, Sunnybrook Health Sciences Centre, Toronto, Ontario, Canada.
- Odette Cancer Program, Sunnybrook Health Sciences Centre, Toronto, Ontario, Canada.
- Department of Radiation Oncology, University of Toronto, Toronto, Ontario, Canada.
- Temerty Centre for AI Research and Education in Medicine, University of Toronto, Toronto, Ontario, Canada.
Abstract
Accurate preoperative identification of cervical lymph node (LN) metastasis is essential for staging and treatment planning in dogs with oral malignancy, yet conventional imaging offers limited diagnostic sensitivity. This retrospective study evaluated whether CT-derived radiomic features, combined with machine learning classifiers, could predict the metastatic status of mandibular and retropharyngeal LNs. Forty-nine dogs with histopathologically confirmed oral malignancy underwent contrast-enhanced CT, and four bilateral cervical LN sites were manually segmented on images resampled to 1.0 mm isotropic spacing, yielding 195 LN observations (18 metastatic, 9.2%). One hundred and seven radiomic features were extracted per LN using PyRadiomics, and variance filtering, Spearman redundancy removal (|ρ| > 0.95), and Mann-Whitney U testing with Benjamini-Hochberg correction were performed within each fold of leave-one-patient-out cross-validation (LOOCV) to eliminate selection-leakage bias. Logistic regression, random forest (RF), support vector machine, and XGBoost were trained with SMOTE applied within training folds only, and two endpoints were evaluated on the same out-of-fold predictions: per-LN (primary) and per-patient (secondary, via maximum-probability aggregation across the four sites). Confidence intervals were derived from 5000-iteration patient-level bootstrap, and sensitivity analyses included repeated StratifiedGroupKFold cross-validation and cluster-robust generalised estimating equation (GEE) inference. At the per-LN level, RF achieved an AUC of 0.649 (95% CI 0.474-0.831; sensitivity 0.444, specificity 0.898) and XGBoost an AUC of 0.631 (0.501-0.772; sensitivity 0.944, specificity 0.379). Patient-level aggregation yielded an RF AUC of 0.613 (0.431-0.786; 9/13 metastatic-positive dogs identified) and an XGBoost AUC of 0.485. Four GLSZM/GLCM texture features were selected in 92%-100% of LOOCV folds and remained significant under GEE (adjusted p ≤ 0.012). CT-derived texture features therefore carry a reproducible, biologically interpretable signal for cervical LN metastasis, but leakage-controlled performance is modest in this proof-of-concept study, and external validation in larger, multi-institutional cohorts is required before clinical translation.