The lack of a consistent MRI radiomic signature, partly due to the multitude of initial feature analyses, limits the widespread clinical application of radiomics for the discrimination of salivary gland tumors (SGTs). This study aimed to identify the optimal radiomics feature category and MRI sequence for characterizing SGTs, which could serve as a step towards obtaining a consensus on a radiomics signature. Preliminary radiomics models were built to discriminate malignant SGTs (n = 34) from benign SGTs (n = 57) on T1-weighted (T1WI), fat-suppressed (FS)-T2WI and contrast-enhanced (CE)-T1WI images using six feature categories. The discrimination performances of these preliminary models were evaluated using 5-fold-cross-validation with 100 repetitions and the area under the receiver operating characteristic curve (AUC). The differences between models’ performances were identified using one-way ANOVA. Results show that the best feature categories were logarithm for T1WI and CE-T1WI and exponential for FS-T2WI, with AUCs of 0.828, 0.754 and 0.819, respectively. These AUCs were higher than the AUCs obtained using all feature categories combined, which were 0.750, 0.707 and 0.774, respectively (p < 0.001). The highest AUC (0.846) was obtained using a combination of T1WI + logarithm and FS-T2WI + exponential features, which reduced the initial features by 94.0% (from 1015 × 3 to 91 × 2). CE-T1WI did not improve performance. Using one feature category rather than all feature categories combined reduced the number of initial features without compromising radiomic performance.

Radiomics analysis can potentially characterize salivary gland tumors (SGTs) on magnetic resonance imaging (MRI). The procedures for radiomics analysis were various, and no consistent performances were reported. This review evaluated the methodologies and performances of studies using radiomics analysis to characterize SGTs on MRI. We systematically reviewed studies published until July 2023, which employed radiomics analysis to characterize SGTs on MRI. In total, 14 of 98 studies were eligible. Each study examined 23–334 benign and 8–56 malignant SGTs. Least absolute shrinkage and selection operator (LASSO) was the most common feature selection method (in eight studies). Eleven studies confirmed the stability of selected features using cross-validation or bootstrap. Nine classifiers were used to build models that achieved area under the curves (AUCs) of 0.74 to 1.00 for characterizing benign and malignant SGTs and 0.80 to 0.96 for characterizing pleomorphic adenomas and Warthin’s tumors. Performances were validated using cross-validation, internal, and external datasets in four, six, and two studies, respectively. No single feature consistently appeared in the final models across the studies. No standardized procedure was used for radiomics analysis in characterizing SGTs on MRIs, and various models were proposed. The need for a standard procedure for radiomics analysis is emphasized.

Wavelet decomposition (WD), widely used in radiomics, redistributes information among derived wavelet components when the input is rotated. This redistribution may alter predictions for the same lesion when scanned at different angles. Despite its potential significance, this vulnerability has frequently been overlooked in radiomic studies while its impact remains poorly understood. Therefore, this study aims to investigate how variations in lesion orientation affect both WD and non-WD radiomic feature values, and subsequently, model performance. We analyzed CT radiomics of primary non-small-cell lung cancer (NSCLC). Prior to feature extraction, we introduced random rotations ranging from 5° to 80° to the tumors. Their effects were quantified by evaluating the percentage difference ( $$\%\Delta$$ % Δ ) between the rotated and unrotated feature values, and validated using Spearman’s rank test. Additionally, radiomics models were trained to discriminate between three histological subtypes of NSCLC using the original features, and then tested on rotated inputs. The correlation between the model accuracies and the degree of rotation was again evaluated using Spearman’s rank test. Four-hundred nineteen NSCLC patients (mean age: 68.1 ± 10.1, 289 men) were evaluated. Significant correlations between feature values and rotations (Spearman’s correlation [CC] magnitude ≥ 0.1, p < .05) were found in 23.7% (176/744) of the WD and 0.5% (5/930) of the non-WD texture features. Significant association between performance and rotation was observed in WD-based models built to discriminate between NSCLC histological subtypes (CC = − 0.44, p < .001) but not in non-WD-based models (CC = 0.03, p = 0.07). Input lesion orientation affects radiomic feature values and model reproducibility. WD features exhibited significantly greater instability to orientation variations compared to non-WD features.