The workflow outlining our approach is illustrated in Figure 1 . In this workflow, the neoplasm volume serves as the region of interest (ROI) from which all radiomics features are computed. The contouring of the ROI was performed manually by experienced Radiation Oncologists, not directly involved in the study, using the Varian Medical System Eclipse software environment. These features underwent a selection process to minimize redundancy and were combined with clinical data. A logistic regression model, optimized via five-repeated 10-fold cross-validation, was then applied. The model’s predictive performance was evaluated using the Area Under the ROC Curve (AUC). All statistical analyses were performed using R programming language, with a significance level (alpha) set at 0.05 for all tests.

Uniformity in voxel sizes is essential for precise and dependable feature calculations in radiomics (49). Given the original CT scan resolutions varied significantly, from 0.3×0.3×0.5 mm to 1.3×1.3×5 mm, we resampled the voxel values to a uniform resolution of 1×1×1mm using interpolation. For this purpose, we employed the Bspline algorithm (50), suitable for smoothly adjusting voxel values. Concurrently, to maintain the binary nature of the tumor masks, we applied the nearest neighbor interpolation algorithm (51, 52), which assigns the mask label of the closest voxel in the original grid to each voxel in the interpolation grid. Another critical step employed was normalization, computed by x − μ σ , where x represents voxel intensity, μ is the average intensity, and σ is the standard deviation of intensity. As demonstrated in a previous study (49), this normalization minimizes variance and boosts feature robustness of radiomic features, especially against different discretization levels, ensured normalized intensities generally spanned from [-3, 3] after outlier exclusion (53, 54). These intensities were then scaled to a range of approximately [-300, 300]. Moreover, to capture detailed textural information, we discretized the intensities within the ROI using a uniform bin width of 5, starting from a normalized minimum HU value of 0. We chose a bin width of 5 to ensure an adequate number of bins (between 1 and 400), allowing for the capture of more detailed textural information (55). This discretization assigns new values to each voxel according to the formula: floor((original intensity)/5) + 1. This method not only suppresses noise but also enhances the robustness of radiomic features by smoothing out minor variations.

Medical images provide insights into the phenotypic traits of neoplasms. These images typically contain data from tens of thousands of voxel intensities per neoplasm, leading to a scenario where the number of features (p) greatly exceeds the sample size (n). In our study, we extracted features from each image set using the PyRadiomics library in Python. According to the Imaging Biomarker Standardization Initiative (IBSI) (56), we extracted features across six categories: shape, first-order statistics, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM) (57), grey level size zone matrix (GLSZM) (58), gray level dependence matrix (GLDM) (59), and neighborhood grey tone difference matrix (NGTDM) (60). Additional calculations were performed on images processed with wavelet, Laplacian of Gaussian (LoG), square, square root, logarithm, exponential, and gradient filters, culminating in 1,092 features.

Logistic Regression Models (LRM) were used in assessing the features of discriminative power. Algorithms are available in the R stats, glmulti package. A logistic model can be mathematically described as follows: log π ( x ) 1 − π ( x ) = β x , where π ( x ) is the probability of recurrence, and β the coefficient vector for the independent variables x (61). The guidelines outlined in (62) recommend that the number of predictors used in fitting LRM should not exceed 10% of the events in the sample. In our cohort, with 10% of events in the training sample, the optimal number of predictors for model fitting should be 2 to 3, as 10% of the total recurrence equals 2.1. Thus, we aimed to limit the final Logistic Regression Model (LRM) to a maximum of 3 degrees of freedom. We used the Best Subset Selection (BSS) Modeling strategy to identify the most effective LRM based on validation performance to achieve this. BSS, known for its efficiency in finding the most parsimonious model, outperforms methods like stepwise selection and Lasso, although its high computational demand is a limitation. For instance, fitting LRMs with 2, 3, or 4 degrees of freedom using 17 radiomics and six clinical features requires BSS to estimate a minimum of 41,262 coefficients, making exhaustive evaluation impractical. Therefore, we reduced the number of input variables before employing BSS.

Radiomics data often faces the challenge of high multicollinearity, where variables are highly correlated, affecting the significance of individual variables in the model. For example, sphericity, minor axis length, and elongation show strong multicollinearity. Multicollinearity can lead to the phenomenon where a variable is not deemed significant when correlated features are also present in the model. Figure 2 uses a color scheme where white represents no correlation, blue represents a perfect negative correlation, and red represents a perfect positive correlation. The heatmap illustrates the correlation coefficients prior to the feature selection process, revealing the initial relationships between features. The heatmap revealed numerous red and blue shades, indicating strong positive and negative correlations, respectively, among the data. There is now a substantial body of research on mitigating multicollinearity, such as Principal Component Analysis (PCA), Sparse PCA (63), and Kernel PCA (KPCA) (64). We employed Recursive Feature Elimination (RFE) (65), an iterative procedure to refine the input data for BSS LRM. Using RFE with repeated 10-fold cross-validation ( Table 1 ), we narrowed down from 1,092 radiomic features to a subset of eight active features ( Figure 3 ). We then fitted degree-2 LRMs, considering all combinations of these eight features across 5,000 data shuffles. The models were trained on 63 samples (80% of the cohort) using 10-fold cross-validation and evaluated based on AUC. The remaining 15 samples (20% of the cohort) were used for prognostic validation. We also used the ROC to visualize the classifiers’ performance. Figure 4 shows that the 3,480 models out of 5,000 data shufflings unveiled the most distinctive factor among the extensive array of 1,092 radiomic features: specifically, Large Dependence Emphasis (LDE) of the Gray Level Dependence Matrix (GLDM) and Long Run Emphasis (LRE) of the Gray Level Run Length Matrix (GLRLM) of the 3D Laplacian of Gaussian (LoG _σ=3) filtered ROIs ( Figures 5 , 6 ). Statistical evidence was quantified by the p-value of the hypotheses’ tests. If the p-value is less than α = 0.05, then the null hypothesis will be rejected in favor of the alternative hypothesis.

We stratified the cohort into high and low LRE and LDE subgroups according to the threshold selected by using respective median. Kaplan-Meier curve analyses were conducted to assess the impacts of the biomarkers on RFS. RFS was defined as the time from surgery to locoregional recurrence.

The primary end-point of the analysis was the time to peritoneal locoregional recurrence. The follow-up duration to peritoneal LR was calculated from the date of surgery to the date when peritoneal LR was diagnosed or to the last follow-up, and information about the survival status and recurrence type was also documented. Finally, a radiomic nomogram was constructed. A model containing both radiomic and clinical factors was also constructed for comparison.

The code used in this study is available at https://doi.org/10.5281/zenodo.10460030.

This retrospective biomarker analysis examines a group of oral cavity squamous cell carcinoma (OSCC) patients who underwent surgical/curative/elective neck/selective neck resection at the institution between 2006 and 2017. The study involved 78 patients, with 21 experiencing locoregional recurrence (LR), while 57 remained disease-free within a 2-year period after the end of the initial treatment course. Demographic and clinicopathological features of patients are detailed in Table 2 . The mean age at the initial surgery was 60, ranging from 30 to 98 years. The median follow-up time for recurrence-free survival (RFS) was 56.2 months. A locoregional recurrence was defined as a positive biopsy in the primary site or the cervical lymphatic region after treatment. We collected six clinical characteristics of interest, including age, gender, tobacco usage, alcohol consumption, T-stage, N-stage, and race. All patients were in the first 2 years of follow-up after surgery. Patients were categorized into four T stages (1, 2, 3, and 4) based on the size and extent of the primary tumor. Smoking and alcohol status were self-reported and coded as 1 for Yes and 2 for No. The missing values for smoking and alcohol status were hard coded as 3 due to their substantial representation within the dataset. Smoking status revealed that 60% of the total cohort were smokers, with this number rising to 72% in the LR subgroup. For alcohol consumption, 40% of the total group reported alcohol use, compared to 48% in the LR subgroup. All patients underwent surgery treatment as primary treatment, along with chemoradiotherapy (CRT) or radiotherapy (RT). The endpoint in this study was 2-year LR status, defined as whether an LR happened within 2 years after curative treatment. Here, the proportion increased in the LR subgroup (17% in the total group, 29% in LR), indicating a higher prevalence of this intermediate stage in the LR subgroup. T3 tumors constituted 17% of the cases. The proportion of T3 tumors increases to 29% in the LR subgroup. T4 tumors, which represent the most advanced stage of tumor size and extent, accounted for 23% of the total cohort. The representation of T4 tumors is notably higher in the LR subgroup, constituting 33%. In total, T3 and T4 stages comprise only 31% of the entire group, in contrast to 62% in the LR subgroup.

A preliminary feature selection algorithm identified 8 radiomic factors of discriminative power in LR depicted in Figure 7 . To ascertain the independence of these two radiomic factors from clinical factors and their potential as clinical alternatives, we investigated their interaction with clinical factors, smoking, alcohol I(ETOH), N stage, and T stage. Logistic regression modeling then incorporated the radiomic features with demographic and clinicopathological characteristics to define the final radiomic risk factors. In this comparative analysis of five logistic regression models in Table 3 , denoted as Models 1 through 5, we have assessed their performance based on a range of statistical metrics. The Akaike Information Criterion (AIC) is employed as a model selection criterion. At the same time, accuracy (ACC), area under the receiver operating characteristic curve (AUC), sensitivities (Sens), and specificities (Spec) are utilized to evaluate the models’ predictive capabilities. It is evident that the 2nd model (clinical-only) exhibits the highest AIC of 80, suggesting a worse fit to the training data (80%) compared to the other models. When considering the training measurements, the model demonstrates the lowest accuracy (0.75) and AUC values (0.46). In regard to the testing measurements on the held-out 20% data, the 2nd model shows the lowest accuracy (0.6) and AUC (0.67) among all models, showcasing its deficiency in generalization. Moreover, the model consistently maintains an unbalanced sensitivity (0.45) and specificity (1), highlighting its inability to make accurate predictions while minimizing false positives and false negatives. These findings collectively underscore the suboptimal performance of the 2nd model and establish it as the least favorable choice when contrasted with the other models in this analysis.

Numerous studies have identified smoking and drinking as risk factors for OSCC patients. Further analysis indicated that including clinical factors didn’t significantly enhance the model’s explanatory power (based on deviance analysis in Table 4 via chi-square test, P=0.43). Table 4 presents the Analysis of Deviance results for eight pairs of nested model comparisons, testing the null hypothesis that additional factors have no effects on outcome. Our analysis yielded robust evidence (via χ2-test, p-values< 0.0001) supporting the significance of adding radiomic factors in each pair. Augmenting radiomics to include Smoke, ETOH, and T (the first pair) decreased the deviance by 27.68, indicating a significantly better fit of the larger model to the data. The larger model’s AUC showed a noteworthy improvement over the smaller model, as depicted in Figure 8 .

Patients were stratified into high- and low-end groups for recurrence-free survival based on the median value of two factors ( Figure 9 ). The Kaplan-Meier curves provide compelling evidence of a significant difference in RFS between the high- and low-end groups (Log-rank p< 0.05). Furthermore, the AUCs of logistic regression, incorporating radiomic factors (0.93), corroborate the significant enhancement in discriminative power when compared to clinical factors-only models ( Figure 8 ). Nomograms were constructed with radiomic and clinical factors respectively in Figure 10 . Notably, the addition of clinical features in the full model demonstrates minimal influence on the predicted RFS probability when compared to the radiomic feature-only model.

Tumor heterogeneity is widely acknowledged as a significant factor associated with tumor progression. The quantification of tumor heterogeneity has assumed a pivotal role in pathological assessments. Radiomic texture analysis presents distinct advantages, such as non-invasiveness and cost-effectiveness, compared to conventional pathological evaluations. Multiple studies (66–69) have underscored the prognostic potential of GLDM (Gray Level Dependence Matrix) and GLRLM (Gray Level Run Length Matrix) features in the evaluation of tumor progression. These features, GLDM and GLRLM, quantify the degree of local variation within an image (70). LRE (Long Run Emphasis) serves as a metric for assessing the distribution of long run lengths, with higher values indicative of longer run lengths and coarser structural textures. Conversely, LDE (Long Dependence Emphasis) quantifies the distribution of large dependencies, with elevated values denoting larger dependencies and more homogeneous textures. In light of the findings presented in Table 5 , it is noteworthy that, when maintaining LDE at a constant value, each unit increment in LRE corresponds to a 79% (1-0.21) decrease in the odds of recurrence as opposed to non-recurrence. Conversely, when keeping LRE at a fixed value, every unit increasing in LDE results in a 5.8% increase in the odds of recurrence. It is essential to recognize that the estimate for the intercept represents the log odds of a patient with hypothetical zero values for LRE and LDE experiencing recurrence, which is calculated to be 0.095. This observation underscores a robust association between these two radiomic factors and the risk of locoregional recurrence.

Figure 11 compares two sets of images processed using the Laplacian of Gaussian (LoG) filter. The first three rows depict LoG-filtered results on an Oral Squamous Cell Carcinoma (OSCC) occurring in the tongue area, with varying sigma (σ) values. The last three rows display results for a non-OSCC area of the tongue. Lower sigma values highlight finer structures, while higher sigma values accentuate larger clusters in the tissue (71). We observed a pattern in the OSCC images: with increasing sigma, there is a reduction in highlights (white regions) in the filtered images, contrary to the non-OSCC images, which maintain a consistent level of highlights. This suggests that squamous cells in OSCC may be more homogeneous than normal cells. Furthermore, we observed circular artifacts in the OSCC images when filtered with larger sigma (σ > 0.4mm), whereas the normal set presents relatively random structures. to normal tissues. These findings support a fundamental histopathological principle: tumor tissues typically exhibit a more anaplastic and infiltrative pattern than normal tissues (72). This occurs because tumor cells grow unregulated and clonally, leading to the loss of normal differentiation and organization characteristic of healthy tissues.