The purpose of this study was to develop a machine learning-based algorithm based on a combination of magnetic resonance imaging (MRI) and color-Doppler ultrasound (CDUS) to characterize lacrimal gland lesions.

All patients with a lacrimal gland lesion who underwent MRI examination and CDUS between 2014 and 2025 were retrospectively included. Thirty-four imaging features were systematically assessed. A machine learning algorithm was trained with repeated nested cross-validation (RNCV) using random forest classifiers. Shapley additive explanations values were used to assess feature contributions. Simplified models using top 5 and top 10 best features were also developed. Diagnostic performance of the models was assessed using area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (PR AUC), balanced accuracy, precision, sensitivity, specificity, Brier score, Matthew's correlation coefficient and F1-score.

One hundred patients (mean age, 49.6 years ± 17.8 [standard deviation] years) with 130 lesions (101 non-epithelial (NEL) and 29 epithelial (EL); 45 malignant) were included. The random forest binary machine learning model yielded 75.9% sensitivity (95% confidence interval [CI]: 39–100), 86.0% specificity (95% CI: 62.4–100), and an AUC of 0.883 (95% CI: 0.692–1.0) for differentiating between malignant and benign lesions and 73.2% sensitivity (95% CI: 33.5–100), 92.9% specificity (95% CI: 69.9–100), and an AUC of 0.93 (95% CI: 0.683–1) for differentiating between EL and NEL. In multiclass analysis (benign NEL, benign EL, malignant NEL and malignant EL), the random forest yielded a macro-averaged AUC of 0.857 (95% CI: 0.722–0.972) for the all-features model. A 5-top features signature comprising apparent diffusion coefficient and resistance index values, echogenicity, age and lesion type (infiltrative vs. well-delineated mass), yielded an AUC of 0.785 (95% CI: 0.641–0.941) to distinguish between the four classes.

A combination of MRI and CDUS features demonstrated high diagnostic performance for characterizing lacrimal gland lesions. A simplified 5-feature signature showed similar diagnostic performance compared to the all-features model and warrants prospective multicenter validation for clinical application.