Pretreatment MRI radiomics for predicting recurrence and metastasis after hypofractionated stereotactic radiosurgery in uveal melanoma

This is the authors' abstract. We don't add key points for this paper.

Purpose To develop and internally evaluate pretreatment MR-based radiomics models for predicting local recurrence and metastasis after hypofractionated stereotactic radiosurgery (SRS) for uveal melanoma. Methods This retrospective study analyzed pretreatment MRI (T1-weighted, T2-weighted, contrast-enhanced T1-weighted) of 88 SRS-treated UM patients. Tumors were segmented on the axial slice with the greatest cross-sectional area, and 854 radiomic features were extracted for each sequence. After inter-observer reproducibility filtering (ICC ≥ 0.80) and elastic-net feature selection, separate Gaussian Naive Bayes models for local recurrence and metastasis were developed using the retained features with a 70:30 train–test split. SMOTE was applied within training folds to address class imbalance. Held-out test set performance was assessed using AUC, F1 score, accuracy, sensitivity, specificity, and predictive values. Results The local-recurrence model achieved an AUC of 0.958, accuracy of 0.889, sensitivity of 1.00 (3/3), and specificity of 0.875 (21/24). The metastasis model showed limited discrimination (AUC, 0.528; accuracy, 0.815; sensitivity, 0.333 [1/3]; specificity, 0.875 [21/24]). Among three-fraction treatments, a total dose >45 Gy was associated with less local recurrence ( p = 0.010), whereas no other clinical variable was significantly associated with either outcome. Conclusion MRI radiomics showed promising discrimination for local recurrence but limited discrimination for metastasis. These findings should be interpreted with caution given the small number of events in the test set and warrant external validation in larger cohorts .