Preliminary study of 3 T-MRI native T1-mapping radiomics in differential diagnosis of non-calcified solid pulmonary nodules/masses

Background Cumulative CT radiation damage was positively correlated with increased tumor risks. Although it has recently been known that non-radiation MRI is alternative for pulmonary imaging. There is little known about the value of MRI T1-mapping in the diagnosis of pulmonary nodules. This article aimed to investigate the value of native T1-mapping-based radiomics features in differential diagnosis of pulmonary lesions. Methods 73 patients underwent 3 T-MRI examination in this prospective study. The 99 pulmonary lesions on native T1-mapping images were segmented twice by one radiologist at indicated time points utilizing the in-house semi-automated software, followed by extraction of radiomics features. The inter-class correlation coefficient (ICC) was used for analyzing intra-observer’s agreement. Dimensionality reduction and feature selection were performed via univariate analysis, and least absolute shrinkage and selection operator (LASSO) analysis. Then, the binary logical regression (LR), support vector machine (SVM) and decision tree classifiers with the input of optimal features were selected for differentiating malignant from benign lesions. The receiver operative characteristics (ROC) curve, area under the curve (AUC), sensitivity, specificity and accuracy were calculated. Z-test was used to compare differences among AUCs. Results 107 features were obtained, of them, 19.5% (n = 21) had relatively good reliability (ICC ≥ 0.6). The remained 5 features (3 GLCM, 1 GLSZM and 1 shape features) by dimensionality reduction were useful. The AUC of LR was 0.82(95%CI: 0.67–0.98), with sensitivity, specificity and accuracy of 70%, 85% and 80%. The AUC of SVM was 0.82(95%CI: 0.67–0.98), with sensitivity, specificity and accuracy of 70, 85 and 80%. The AUC of decision tree was 0.69(95%CI: 0.49–0.87), with sensitivity, specificity and accuracy of 50, 85 and 73.3%. Conclusions The LR and SVM models using native T1-mapping-based radiomics features can differentiate pulmonary malignant from benign lesions, especially for uncertain nodules requiring long-term follow-ups.


Introduction
Lung cancer is the leading cause of cancer death in men aged ≥ 40 years and women aged ≥ 60 years, causing far more deaths than breast cancer, prostate cancer, etc. Although tobacco control and improved treatment methods have reduced the mortality rate of lung cancer, it is estimated that 69,410 men (about 22%) and 62,470 women (about 22%) in the United States would still die of lung cancer in 2021, ranking the first among cancer deaths, according to the latest report [1].
Approximately 3.6-24.2% of screening LDCT scans were classified as indeterminate or positive [2], while the majority (84-95%) of the positive scans were proven false-positive [3,4]. In addition, indeterminate results require at least one follow-up examination after few months, especially for nodules with size of 4-8 mm, which need serial recurrent CT scans at interval of 3, 6 and 9 months [5]. Cumulative radiation dose from repeated CT scans should be taken in account, either physical damage or psychological stress [6,7], in particular, an extreme threat to children, the pregnant women, and those with low immunity. Magnetic resonance imaging (MRI), safer to those special people, is an alternative tool for chest lesions, with advantages of non-radiation, multi-parameters and functional measurements. Diffusion-weighted imaging (DWI), intra-voxel incoherent motion imaging (IVIM) and dynamic contrast enhanced MRI (DCE-MRI) are useful for differential diagnosis of pulmonary nodules [8][9][10][11][12]. When differentiating malignancy from benign nodules, the pooled sensitivity and specificity using DWI reached 80-88 and 89-93%, respectively [13,14]. However, those MRI scans are limited by unsatisfactory repeat-ability caused by measurement bias and frequently motion artifacts, especially for nodules less than 20 mm [15,16].
Radiomics features, for instance, first-order features and texture features, extracted from medical imaging are able to present the inherent heterogeneity of lesions, like coarseness, which are useful in the differential diagnosis of suspicious pulmonary nodules [17][18][19][20][21]. The 4-radiomic features of short axis, contour, concavity, and texture had an validation test AUROC of 0.8 (accuracy = 74.3%, sensitivity = 66.7%, specificity = 75.6%) in predicting malignancy in primary nodules [22]. The accuracy using the 4 signatures, including Laws_LSL_min, Laws_SLL_energy, Laws_SSL_skewness and Laws_EEL_ uniformity, in benign or malignant classification was 84%, with the sensitivity of 92.9% and the specificity of 72.7% [23]. However, the majority of researches on pulmonary nodules radiomics features are mainly rely on CT and PET/CT modalities, and there are few reports on MRI. Additionally, sequences for instance, DWI, with obvious deformation and local drift, are unfavorable for the stability and repeat-ability of radiomics features. Native T1-mapping, obtained in one breath-holding scan, can quantify the T1-value of pixels or tissue, compared to conventional T1-weight imaging, which could be a candidate for differentiating malignant from benign lesions.
Our previous trial [24] confirmed that native T1-mapping was comparable to CT for evaluating nodules morphology. Moreover, the native T1-value is potentially useful in discriminating malignancy or tuberculosis from non-tuberculosis benign lesions. However, the value of radiomics signatures based on native T1-mapping imaging in differential diagnosis of pulmonary lesions are still not well established.
In summary, we assume that radiomics features based on native T1-mapping were capable of distinguishing the malignancy from benign lesions. Thus, this prospective study intended to uncover the underlying radiomics signatures of native T1-mapping and further investigated the value of native T1-mapping in aiding to differentiate malignant from benign lesions.

MRI scan
All patients underwent MRI examination using 3 T whole-body MR scanner (MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany) with an 18-element body wrap coil. All parameters were listed as follows: axial T1-

Nodule segmentation
The region of interest (ROI) was segmented across all of the two-dimensional T1-mapping sections of the lesions with an in-house semi-automatic hand-annotation tool in axial view using open-source software (Multi-labe, version 1.1; Shanghai Key laboratory of Magnetic Resonance, East China Normal University, China), as shown in the Fig. 2. The radiologist was blinded to pathological diagnosis but was provided with clinical information such as age, and was also given the option to vary the window and level setting within this software to efficiently annotate the nodule. Manually and semi-automatic methods were optional for nodules annotation. The nodules or masses were annotated by the same radiologist again after 6 months for evaluating intra-observer's agreement.

Feature extraction and grouping
A total of 107 radiomics features were extracted from the annotated mask with an open-source platform PyRadiomics (https:// pyrad iomics. readt hedocs. io/ en/ lates t/) [25], which enables the processing and extraction of radiomic features from medical image data and is implemented in Python. The acquired images features comprised the first-order statistics (18 features

Dimensionality reduction and radiomics feature selection
In order to avoid the over-fitting, a three-step dimensionality reduction was conducted in turns. Firstly, ICC was used for intra-observer's agreement of radiomics features. ICC-value greater than 0.60 was considered a good agreement [26]. Secondly, t-test and wilcox-test were used for univariate analysis. P-value less than 0.05 was considered as statistical significance. Then, LASSO analysis was used for dimensionality and feature selection. tenfold cross-validation was used to reduce the over-fitting and enhance the robustness of model performance.

Radiomics signature and models
Three classical machine methods, binary LR, SVM and decision tree were used to select optimal features and develop the model for that differential diagnosis. Multivariable binary logistic regression with backward stepwise selection was used to build a linear classifier. Then, the SVM with a Gaussian kernel was used to build a nonlinear classifier. Optimizing the parameters of the SVM kernel function were realized by tenfold cross-validation, which was capable of selecting the best performing signature. The decision tree is a simple linear classifier with a sequence of questions. tenfold cross-validation and reduction of branches were applied to avoid over-fitting.

Statistical analysis
All statistical analysis were conducted using R software (version, 4.0.3, http://r-proje ct. org). The Shapiro-Wilk test was used to assess the normality of distributions, and the homogeneity of variance was tested using Bartlett's test. The assessment of the optimal radiomics signature and diagnostic performance mainly relied on ROC. Accuracy, sensitivity, and specificity, were also calculated at the maximum of Youden's index. Each radiomics signature was also assessed according to all of the metrics for the validation cohort. Z-test was used for comparisons of diagnostic performance of inter-models. P-value less than 0.5 was considered as statistical significance.

ICCs of radiomics features
Results of the intra-observer's agreement manifested that 19.5% of all 107 features had a relatively Fig. 3 The radiomics features extracted from the native T1-mapping satisfactory agreement with ICCs of greater than 0.6 (mean ICC = 0.41, range from 0 to 0.87) ( Fig. 4; Table 2). Collected for further analysis.

Dimensionality reduction and radiomics feature selection
Univariate analysis results manifested that 21 features were preserved, which were further used for lassoanalysis. 5 features were selected by Lasso-analysis, with best tuned regularization parameter λ of 0.062 under the 1-SE criteria found by tenfold cross-validation. The remaining 5 features were glcm_MCC, glcm_DifferenceVariance, glcm_Imc1, glszm_GrayLev-elNonUniformity and shape_MinorAxisLength, respectively (Table 3).

Discussion
In this present study, we developed diagnostic models based on native T1-mapping images to differentiate malignant from benign lesions. According to our results, the SVM and LR classifiers both had satisfactory AUC of 0.82, with sensitivity of 70%, specificity of 85% and accuracy of 80%. The SVM classifier is a powerful tool to analyze data with large number of predictors and limited sample sizes, especially when handling binary outcomes [27]. This is consistence with previous studies [28]. In addition, prior analysis also showed that the LR model based on MRI-radiomics had best performance compared to other classifiers like decision tree, k-nearest neighbor, and XGBoost, etc. [29,30]. In order to minimize the risk of modeling over-fitting and bias, the LASSO and tenfold cross-validation were used for feature selection and models constructing. In general, our results also manifested that the SVM and LR classifiers, superior to decision tree, are suitable for diagnostic models in indeterminate pulmonary lesions. At present, the diagnostic work-up of lung cancer using imaging radiomics mainly rely on CT, PET-CT and MRI, of which, CT is the most used. Recently, Gillies, et al. had AUC of 0.80 and 0.85 in discrimination of lung cancer, using size and shape features, non-size based features respectively [31]. Garau et al. found that the handcrafted LDCT radiomics features model (LASSO + SVM) had a higher AUC than the LungRADS clinical model (0.86 vs. 0.76) in the external validation, when identify malignancy [32]. Radiomics features are also valuable in histopathological classification, prognostication, treatment response, and gene mutation, etc. In this study, the radiomics features extracted from native T1-mapping also had clinically significant AUC of 0.82 for malignancy identification. And MRI, being free of radiation dose, is a relatively ideal follow-up inspection tool, especially for women, pregnant women, and people with low immunity. In clinical practice, it is little known about the radiomics features based on MRI modality, mainly T2WI and DWI sequences. Using RFE with SVM, the joint model of T1WI, T2WI, and ADC showed the highest performance with AUC of 0.88 in classification of pulmonary lesions [33]. Besides, radiomics signatures extracted from ADC, DWI, T2WI can be used for predicting EGFR mutation in patients with lung adenocarcinoma [28]. Compared to DWI or IVIM [33], native T1-mapping, obtained in a single breath-holding, had almost no artifacts, deformation and location shift, which is save-timing and more favorable to ensure stability and repeatability of radiomics features. And as far as we know, our team was the first to investigate the value of native T1-mapping radiomics features in differentiating malignant from benign lesions.
Our study showed that the textures features including glcm_MCC, glcm_DifferenceVariance, glcm_Imc1, and glszm_GrayLevelNonUniformity in malignancy were higher than benign lesions, which indicates more significant heterogeneity in malignant lesions. Additionally, the MinorAxisLength in this study had statistically significance in malignancy identification, which could be explained that the lung cancer are usually larger in size than benign lesions. Interestingly, unlike previous report [33], the surface area to volume ratio (SAVR) of lesions was useless in present study. The SAVR presents the degree of a sphere-like shape. We assume that the thickness of the scan sequence has a great influence on the morphology of the lesion. The thinner thickness is more helpful to the value of shape features. Another thing worth mentioning is that there were more benign lesions than malignancy in our samples, which is line with the clinical scenario.
Generally, there are several limitations to this study. Firstly, additional cases were needed for constructing radiomics-based model, which is very important to strengthen the reliability and improve the diagnostic performance. Secondly, the segment of lesions was conducted by one radiologist at a different point in time. The inter-observer's agreement was absent. Most importantly, this is a single-center experiment and performed on one scanner. Multicenter validation and multi-scanners' variations still need great effort. Moreover, the effect of image reconstruction and scanners on variability of MRIbased radiomic features needed further investigation. In addition, native T1-mapping failed to identify the mass and secondary obstructive pneumonia. We assumed that post-contrast enhancement T1-mapping would produce better contrast between tumor and background.

Conclusion
Texture features based on native T1mapping are useful for differentiating pulmonary malignant from benign lesions. The optimal SVM and LR model using 5 texture features acquired the AUC of 0.82, the sensitivity of 70.0%, specificity of 85.0% and accuracy of 80.0%, respectively, in the validation set. Native T1-mapping could be an compensatory tool for the management of pulmonary nodules, especially for those who need long-time follow-ups.