In the present study we developed and validated an ADC-based radiomics signature for predicting histologic grade in SCC of the tongue and MF. Based on a significant difference in the radiomics score, the radiomics signature could successfully differentiate between low- and high-grade tumors.
During pretreatment evaluation, the histologic grade of SCC of the tongue and MF has long been considered an important prognostic factor following the TNM stage independently or with other biomarkers[7, 16]. Invasive biopsy is the “gold standard” for pretreatment evaluation of histologic grade, but is often limited due to sample bias. As a non-invasive approach for characterizing tumors comprehensively, ADC-based radiomics has demonstrated potential for prediction of the histologic grade of giloma, cervical cancer, and bladder cancer. Therefore, we investigated the predictive ability of radiomics features from ADC maps for the histologic grade of SCCs of the tongue and MF.
In this study, a supervised LASSO method was used to construct a radiomics signature for predicting histologic grade, which could provide comprehensive information about tumor heterogeneity. This method is designed to avoid overfitting for analyzing large numbers of radiomics features with a relatively small sample size, and has been used widely in radiomics research [21, 22]. 231 candidate features associated with histologic grade were reduced to six potential predictors to build an ADC-based radiomics signature. The signature showed a significant difference between low- and high-grade tumors in both cohorts, and performed well for predicting histologic grade, with an AUC of 0.82 for the training cohort and 0.78 for the testing cohort. Ahn et al. reported that the mean ADC from a high b-value (2000 s/mm2) DWI had an accuracy of 0.70 for distinguishing low- and high-grade head-and-neck SCCs, but was not validated. In our work, on the basis of the cutoff values from the training cohort, the radiomics signature could differentiate between low- and high-grade tumors in the testing cohort with an accuracy of 0.78, which suggested that the ADC-based radiomics signature outperformed the measurement of mean ADC from DWI with high b value. Given the comparable proportions of low- and high-grade tumors in both cohorts, the similar predictive performance implied that the radiomics signature was robust. If our study data can be reproduced from other centers, our results would likely suggest that an ADC-based radiomics signature could be used for computer-aided grading of SCCs of the tongue and MF.
In agreement with previous studies [6, 8], mean ADC on original ADC maps with b = 0 and 1000 s/mm2 was found to be not significantly different between low- and high-grade tumors. The more prevalent microscopic necrotic areas and peritumoral edema with an increased ADC value and absence of keratinization of cells known to hinder water diffusion would explain (at least in part) the higher-than-expected ADC in high-grade tumors . Of the optimal radiomics features, positive skewness on original ADC maps was demonstrated in low- and high-grade tumors, and was significantly greater in the latter. Skewness indicates the asymmetry of the histogram distribution, with a higher positive skew denoting that the voxel values cluster toward the lower end of the histogram . Studies have reported significantly higher ADC skewness in high- than low-stage renal cell carcinomas , as well as malignant compared with benign endometrial tumors . Our observation of higher ADC skewness within high-grade SCCs of the tongue and MF reflected a predominance of lower ADC values resulting from neoplasia-related cellularity. In addition, Histogram_Entropy, GLRLM_RunEntropy, GLCM_InverseVariance, and GLCM_Correlation of ADC maps after wavelet decomposition were significantly higher in high-grade than in low-grade tumors. The histogram, GLCM, and GLRLM are typical first-, second- and high-order descriptions of the pixel distribution within a ROI, and can characterize the global, local, and regional heterogeneity of tumors on different scales . Presumably, these findings reflect the fact that high-grade tumors have greater intratumoral heterogeneity than low-grade tumors resulting from increased hypoxic voids, necrosis, and edema within the tumor.
The present study had four main limitations. First, this single-center study was relatively small, and multicenter researches with larger patient cohorts are needed to confirm our findings. Second, Grade-II and Grade-III tumors were not analyzed separately because of the small number of samples. Third, delineating the whole tumor volume was challenging because some tumors were inﬁltrative with indistinct borders. Fourth, the possibility of type-I errors >0.05 could not be avoided without adoption of multiple-testing correction. However, given the small sample size and exploratory purpose of our preliminary study, multiple-testing correction would not have been appropriate.