Texture Analysis of Fat-Suppressed T2-Weighted Magnetic Resonance Imaging and Use of Machine Learning to Discriminate Nasal and Paranasal Sinus Small Round Malignant Cell Tumors

Objective We used texture analysis and machine learning (ML) to classify small round cell malignant tumors (SRCMTs) and Non-SRCMTs of nasal and paranasal sinus on fat-suppressed T2 weighted imaging (Fs-T2WI). Materials Preoperative MRI scans of 164 patients from 1 January 2018 to 1 January 2021 diagnosed with SRCMTs and Non-SRCMTs were included in this study. A total of 271 features were extracted from each regions of interest. Datasets were randomly divided into two sets, including a training set (∼70%) and a test set (∼30%). The Pearson correlation coefficient (PCC) and principal component analysis (PCA) methods were performed to reduce dimensions, and the Analysis of Variance (ANOVA), Kruskal-Wallis (KW), and Recursive Feature Elimination (RFE) and Relief were performed for feature selections. Classifications were performed using 10 ML classifiers. Results were evaluated using a leave one out cross-validation analysis. Results We compared the AUC of all pipelines on the validation dataset with FeAture Explorer (FAE) software. The pipeline using a PCC dimension reduction, relief feature selection, and gaussian process (GP) classifier yielded the highest area under the curve (AUC) using 15 features. When the “one-standard error” rule was used, FAE also produced a simpler model with 13 features, including S(5,-5)SumAverg, S(3,0)InvDfMom, Skewness, WavEnHL_s-3, Horzl_GlevNonU, Horzl_RLNonUni, 135dr_GlevNonU, WavEnLL_s-3, Teta4, Teta2, S(5,5)DifVarnc, Perc.01%, and WavEnLH_s-2. The AUCs of the training/validation/test datasets were 1.000/0.965/0.979, and the accuracies, sensitivities, and specificities were 0.890, 0.880, and 0.920, respectively. The best algorithm was GP whose AUCs of the training/validation/test datasets by the two-dimensional reduction methods and four feature selection methods were greater than approximately 0.800. Especially, the AUCs of different datasets were greater than approximately 0.900 using the PCC, RFE/Relief, and GP algorithms. Conclusions We demonstrated the feasibility of combining artificial intelligence and the radiomics from Fs-T2WI to differentially diagnose SRCMTs and Non-SRCMTs. This non-invasive approach could be very promising in clinical oncology.


INTRODUCTION
Malignant tumors in the nasal and paranasal sinuses are rare, comprise less than 1% of all malignancies and about 3% of head and neck malignancies (1,2), including small round cell malignant tumors (SRCMTs) and non-SRCMTs. SRCMTs form a specific group of malignancies in the nasal and paranasal sinuses based on neuroectodermal, soft tissue, and hematopoietic differentiation, such as seen in rhabdomyosarcoma (RMS), malignant melanoma (MM), olfactory neuroblastoma (ONB), neuroendocrine carcinoma (NEC), and lymphoma. In contrast, non-SRCMTs form another common group of malignant tumors in the nasal and paranasal sinuses based on epithelial differentiation, including squamous cell carcinomas (SCCs) and adenoid cystic carcinomas (ACCs) (3). The distinction between these two groups is crucial as tumors are variably managed with radiation, chemotherapy, conservative medical therapy, local surgery, exenterative surgery, and multimodal therapy, indicating that therapeutic decisions, surgical planning, and prognoses are very different for each tumor type (4).
Conventional magnetic resonance imaging (MRI) has limitations of its own when differentiating between SRCMTs and Non-SRCMTs. Under the circumstances, as texture analysis (TA) techniques, by using mathematically defined features, can analyze pixel distributions, intensities and dependencies, it can provide a wealth of information beyond what can be seen with the human eye and thus can be used to characterize SRCMTs and Non-SRCMTs, quantitatively (5). Other sequences, such as the apparent diffusion coefficient, have been used to discriminate benign and malignant nasal and paranasal sinus lesions or different histopathologic types of sinonasal malignancies (6)(7)(8)(9)(10)(11). However, less attention has been given to the application of TA for fat-suppressed T2-weighted MR images (Fs-T2WI) collected as part of routine clinical practice.
As a branch of artificial intelligence, machine learning (ML) includes various algorithms that can enhance diagnosis, treatments and follow-up results in neuro-oncology medicine by analyzing huge complex datasets (12,13). More importantly, not depending on user experience, ML is more objective than other conventional analyses and has good repeatability. To our knowledge, no studies using TA and ML for differentiating sinonasal SRCMTs from non-SRCMTs have been reported. To bridge this gap, this retrospective study was intended to evaluate the potential value of the ML-based Fs-T2WI texture analysis for distinguishing SRCMTs from non-SRCMTs. To achieve the optimal predictive ability and clinical utility, we compared two-dimensional reduction, four feature selection methods and ten ML algorithms.

Patients
We used the surgical pathology database from January 1, 2018, to January 1, 2021, in our hospital. Exclusion criteria were (1) patients who received treatments before MRI scans and (2) inadequate image quality. All methods were performed in accordance with the relevant guidelines and regulations, and the informed consent requirement was waived. This retrospective study was approved by the Institutional Ethics Review Committee of our hospital.

Extraction of Textural Features
MaZda software (version 4.7, The Technical University of Lodz, Institute of Electronics, http://www.eletel.p.lodz.pl/mazda/) was used for the analyses. We applied the limitation of dynamics to m± 3d(m: mean grey-level value, d: standard deviation) (14) to achieve reliable results for the MRI texture classifications. Regions of interest (ROIs) on the Fs-T2WI images of the largest layer were selected. Two physicians delineated ROIs manually along the edge of the lesion and filled the lesion in with a red marker, excluding the various necrotic and cystic regions. In total, 271 features were extracted for each ROI. The number of radiomics features based on feature classes were as follows and shown in Table 1: (i) 9 histogram features based on the number of pixel counts in the image that possessed a certain grey-level value (15); (ii) 220 grey-level co-occurrence matrix (GLCM) features based on the extraction of statistical information about the distribution of pixel pairs (16); (iii) 20 grey-level run-length matrix (GLRLM) features based on searching the image for runs that have the same grey-level values in a pre-defined direction (17); (iv) a 5 auto-regressive model (ARM) based on the weights associated with four neighboring pixels and the variance of the minimized prediction error; (v) 12 wavelet transform (WAV) features on texture frequency components extracted from the energies computed within the channels (18); and (vi) 5 absolute gradient statistics (AGS) features based on smooth or steep variations, resulting in low or high gradient values (15). Multiple GLRLMs were computed along the 0°, 45°, 90°, 135°, and z-axis directions, and 1, 2, 3, and 4 pixels. Multiple GLCMs were computed along four different angles (horizontal, vertical, diagonal 45°, and diagonal 135°).

Feature Selections
Computer-generated random datasets were used to assign 70% of datasets to the training set and 30% of the datasets to the independent test set. FeAture Explorer software (FAE, V 0.3.6) software on Python (3.7.6) (https://github.com/salan668/FAE) was used. Firstly, the synthetic minority oversampling technique (SMOTE) was used to balance the training dataset. This method works by taking each minority class sample and introducing synthetic examples along the line segments joining any/all of the k minority class nearest neighbors. The neighboring points were randomly chosen depending on the amount of over-sampling required. Secondly, we normalized the dataset by Z-score Normalization, which subtracts the mean value and divides the standard deviation for each feature. Lastly, we used a Pearson Correlation Coefficient (PCC) and principal component analysis (PCA) to reduce the dimensions. PCC is used for each pair of two features to reduce the row space dimensions of the feature matrix (19). If the PCC was larger than 0.99, one of them was randomly removed. PCA is an unsupervised feature reduction technique that explains the variance-covariance structure of a set of variables through linear combinations. Analysis of Variance (ANOVA) and Kruskal-Wallis (KW) and Recursive Feature Elimination (RFE) and Relief (20) were used for the feature selection. ANOVA was a common analytic method to explore the significant features corresponding to the labels. The KW is a non-parametric version of ANOVA, which hypothesizes that the population median of all groups is equal. The relief selects the sub-data set and finds the relative features according to label recursivity. The goal of the RFE is to select features based on a classifier by recursively considering a smaller set of features. The feature number range was set from 1 to 20.

Evaluations
The results were evaluated using leave-one-out cross-validation (LOOCV). Using LOOCV, learning sets were created by taking all samples but one, which was used as the validation set. The accuracy, sensitivity, and specificity were also calculated at a cutoff value that maximized the value of the Youden index. The area under the receiver operator characteristics curve (AUC) for the classification of results was calculated for each tested condition.

RESULTS
Of the 171 consecutive patients with a pathologic diagnosis of SRCMTs or Non-SRCMTs over a 2-year period from January 2018 until January 2021, seven were excluded for poor MRI image quality, and 164 patients were finally selected for the study. There were 70 patients with SRCMTs and 94 patients with Non-SRCMTs; RMS (n=16), lymphoma(n=18), MM (n=14), NEC (n=14), ONB (n=8), SCC (n=66), and ACC (n=28). There were 94 males and 70 females in the entire cohort. The mean age of the patients was 55.22 years with a range of 13 to 87 years. After removing invalid cases automatically with FAE, 162 cases were included with 68 SRCMTs and 94 Non-SRCMTs. We assigned 70% of the datasets to the training set (114 patients with 48 SRCMTs and 66 Non-SRCMTs) and 30% of datasets to the independent test set (48 patients with 20 SRCMTs and 28 Non-SRCMTs).
The SMOTE technique was used to automatically create 18 synthetic SRCMTs samples in the training set by operating in the feature space. We compared the AUC of all pipelines on the validation dataset with FAE. The pipeline using PCC dimension reduction, Relief feature selection, and a GP classifier yielded the

DISCUSSION
This study investigated the potential value of the Fs-T2WI texture analysis of maximum tumor solid components for distinguishing SRCMTs from non-SRCMTs with ML. The key findings were as follows: (1) The pipeline using PCC dimension reduction, relief feature selection, and GP classifier yielded the highest AUC. (2) The best algorithm was GP whose AUCs of the training/validation/test datasets by the two-dimensional reduction methods and four feature selection methods were greater than approximately 0.800. Especially, the AUCs of different datasets were more than about 0.900 using the PCC, RFE/Relief, and GP algorithm. (3) TA with ML appears to be most helpful in tumor differentiation using standard Fs-T2WI routinely acquired with a high accuracy of 0.89. Radiomics data contain first-, second-, and higher-order statistics (22). First-order statistics are described as the   Second-order statistics describe texture features. Specifically, they describe statistical interrelationships between voxels with similar or dissimilar contrast values, such as the co-occurrence matrix, which is calculated from the intensities of pixel pairing, with the spatial relationship of pixel pairing defined. Higherorder statistics impose filter grids on images to extract repetitive or nonrepetitive patterns, such as wavelets, which are data on texture frequency components extracted from the energies computed within the channels. In 2015, Fujima et al. (23) assessed the utility of the histogram analysis on tumor blood flow (TBF) obtained with pseudo-continuous arterial spin labeling to differentiate SCC and lymphoma in nasal or sinonasal cavities, achieving an accuracy of 0.87 for the mean TBF, the coefficient of variation, and kurtosis. In 2019, this group (24) also applied histograms and TAs for Fs-T2WI to differentiate SCC and lymphoma in the head and neck and found that the relative mean signal, contrast, and homogeneity could be useful. Another study (25) (34), for another example, analysed between active contour and Otsu thresholding segmentation algorithms in segmenting brain tumor MRI and pleura diseases CT, respectively. In addition, Hussein et al. (35) proposed a new Viola-Jones model for the segmentation of ovarian and breast ultrasound images. Artificial neural networks and SVM have been tried for division of nasopharyngeal carcinoma, respectively (36,37). The accuracy and consistency of the tumor delineation plays an important role in differential diagnosis. However, most of the tumors in the site of the nasal cavity and paranasal sinus adjacent to the air in Fs-T2 maps are without edema areas. Thus, we chose manual segmentation by experts not automated or semi-automatic segmentation to determine the boundary.
There were some limitations. First, as the SRCMTs studied were of various histologic types, subgroup analyses in more details should be performed in future studies after obtaining a larger sample size and a careful consideration of the study groups. Second, our model used manually delineated ROIs performed along the edge of the lesion. Segmenting precise tumor regions is the focus of future work. In our further studies, we will propose a multiparametric MRI investigation including ADC, T2-weighted MRI and dynamic contrastenhanced MRI involving early and delayed phases to generate a robust model to differentially diagnose SRCMTs and Non-SRCMTs by segmenting precisely three-dimensional tumor regions in a larger sample.

CONCLUSIONS
We demonstrated the feasibility of combining artificial intelligence and radiomics using Fs-T2WI in the differential diagnosis of SRCMTs and Non-SRCMTs. This approach could be a very promising non-invasive method in clinical oncology.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the Institutional Ethics Review Committee of West China Hospital. Written informed consent from the participants' legal guardian/next of kin was not required to participate in this study in accordance with the national legislation and the institutional requirements. Written informed consent was not obtained from the individual(s), nor the minor(s)' legal guardian/next of kin, for the publication of any potentially identifiable images or data included in this article.

AUTHOR CONTRIBUTIONS
CC, YQ and JC initiated this study, participated in its design, performed study selection, data extraction, and data analysis. CC drafted the manuscript. FG supervised all aspects of the study. XZ revised the language. All authors contributed to the article and approved the submitted version.

FUNDING
This study was supported by the National Natural Science Foundation of China (no. 81930046, 81771800, and 81829003).