|Year : 2022 | Volume
| Issue : 2 | Page : 37-46
Magnetic resonance imaging-based radiomics analysis for the assessment of hepatic alveolar echinococcosis biological activity: A preliminary study
Zhoulin Miao1, Ren Bo2, Yuwei Xia3, Wenya Liu4
1 Department of Radiology, Qingdao Municipal Hospital, Qingdao, China
2 First Affiliated Hospital of Xinjiang Medical University, Urumqi, China
3 HuiYing Medical Technology Co., Ltd., Urumqi, China
4 Imaging Center, First Affiliated Hospital of Xinjiang Medical University, Urumqi, China
|Date of Submission||21-Mar-2022|
|Date of Acceptance||25-Jun-2022|
|Date of Web Publication||8-Nov-2022|
Imaging Center, First Affiliated Hospital of Xinjiang Medical University, Urumqi Xinjiang
Source of Support: None, Conflict of Interest: None
OBJECTIVE: The objective of this study was to develop and evaluate predictive models based on a combination of T2-weighted images (T2WI) and different machine learning algorithms, and to explore the value of hepatic alveolar echinococcosis (HAE) activity assessment by magnetic resonance imaging (MRI) radiomics.
MATERIALS AND METHODS: This retrospective study included 136 patients diagnosed with HAE at the First Affiliated Hospital of Xinjiang Medical University between 2012 and 2020. All subjects underwent MRI and positron emission tomography–computed tomography (PET-CT) before surgery. Taking the PET-CT examination results as the reference standard, patients were divided into active (90 cases) and inactive groups (46 cases). The volume of interest of the lesion was manually delineated on T2WI, and quantitative radiomics features were extracted. Synthetic Minority Oversampling Technology was used to balance the number of patients in the categories. To control for redundancy, the least absolute shrinkage and selection operator was used for feature screening after normalization, and ten optimal features were obtained based on correlation coefficient screening. Three machine learning classifiers were trained using five-fold cross-validation and their performance was compared to establish an optimal HAE activity assessment model. The performance of the classifier was evaluated by area under the receiver operating characteristics curve (AUC), sensitivity, specificity, and accuracy (ACC). The ten optimal features selected from each fold were combined using three machine learning algorithms: logistic regression, multilayer perceptron (MLP), and support vector machine, to establish an HAE activity prediction model.
RESULTS: The three machine learning classifiers all showed good prediction performance with a mean AUC on the test set of more than 0.80, and the MLP showing the best performance (AUC = 0.830 ± 0.053, ACC = 0.817, sensitivity = 0.822, and specificity = 0.811).
CONCLUSION: HAE activity can be accurately evaluated by a radiomics method using a combination of quantitative T2WI features and machine learning.
Keywords: Activity evaluation, hepatic alveolar echinococcosis, magnetic resonance imaging, positron emission computed tomography, radiomics
|How to cite this article:|
Miao Z, Bo R, Xia Y, Liu W. Magnetic resonance imaging-based radiomics analysis for the assessment of hepatic alveolar echinococcosis biological activity: A preliminary study. Radiol Infect Dis 2022;9:37-46
|How to cite this URL:|
Miao Z, Bo R, Xia Y, Liu W. Magnetic resonance imaging-based radiomics analysis for the assessment of hepatic alveolar echinococcosis biological activity: A preliminary study. Radiol Infect Dis [serial online] 2022 [cited 2023 Mar 23];9:37-46. Available from: http://www.ridiseases.org/text.asp?2022/9/2/37/360503
| Introduction|| |
Hepatic alveolar echinococcosis (HAE) is a rare parasitic disease caused by Echinococcus multilocularis invasion of the liver, and the disease has a malignant growth pattern. The prevalence of HAE has a certain regional nature, occurring mainly in provinces and regions with developed animal husbandry. Patterns of its growth are exophytic and invasive, with it metastasizing through the blood and lymphatic system, in which situation it is known as “parasite cancer” and is extremely harmful to the human body. According to the relevant guidelines, if a patient has the following conditions: (1) the systemic condition cannot tolerate surgery; (2) advanced multi-organ vesicular echinococcosis has removed the opportunity for radical resection and liver transplantation; and (3) the patient is waiting for liver transplantation; adjuvant treatment before and after surgery with oral albendazole should be followed per the doctor's instructions, and regular follow-up should be set to assess the activity of lesions to guide continuing medication. Therefore, the accurate judgment of lesion activity is important for guiding the treatment plan. Positron emission tomography–computed tomography (PET-CT) has been used as the reference standard to evaluate lesion activity, but it has certain limitations in terms of economic efficiency and equipment availability. Therefore, it is important to find another method to accurately assess HAE activity because this assessment is important for the selection of treatment and determination of the prognosis. In 2012, Lambin et al. proposed the concept of radiomics analysis based on the heterogeneity characteristics of solid tumors. This analysis can extract a large number of high-throughput image data features and excavate larger quantities of more reliable tumor heterogeneity information than doctors can identify with the naked eye. Radiomics effectively converts medical images into high-dimensional recognizable feature spaces, facilitating statistical analysis of the resulting feature spaces to build models with diagnostic, prognostic, or predictive value, thereby providing valuable information for personalized diagnosis and treatment. As part of the emerging trend for combining medical treatment and artificial intelligence, radiomics is gradually being applied to the diagnosis and treatment of lesions, and magnetic resonance imaging (MRI) radiomics has achieved promising results in clinical applications.,,, The objective of this study was to apply radiomics methods to MRI of patients with HAE to construct an imaging feature-based HAE activity prediction model, thereby providing a tool for evaluating the biological activity of lesions. Introducing this new AI technology into HAE diagnosis and treatment forms pioneering work.
| Data and Methods|| |
This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (approval number: 20190225-108), and a waiver for informed patient consent was granted. A total of 156 cases of HAE were diagnosed in the First Affiliated Hospital of Xinjiang Medical University from 2012 to 2020, of which 136 patients were finally enrolled, with 89 being confirmed by pathology. Of the 136 enrolled patients, 90 had active HAE and 46 had inactive HAE. The inclusion criteria were as follows: (1) patients with HAE confirmed by imaging or pathology; (2) abdominal MRI scans and PET-CT scans available for assessment on the picture archiving and communication system; and (3) primary isolated lesion. The exclusion criteria were as follows: (1) image quality not meeting diagnostic requirements; (2) image data mismatch; (3) patients with extensive intrahepatic fibrosis, nodules, or old lesions; and (4) cases with previously diagnosed HAE and surgical intervention.
Magnetic resonance imaging
MRI was performed using 1.5-T (Siemens Avanto, Germany) and 3.0-T (Siemens Skyra, Germany) and 1.5-T (Philips Achieva, The Netherlands) scanners. Body coils were used and patients were scanned in the supine position. The scans ranged from the top of the diaphragm to the lower border of the liver. T1-weighted imaging (TIWI) and T2WI were acquired with 10–28 slices, slice thickness of 5–7 mm, and slice spacing of 1–1.6 mm.
Positron emission tomography–computed tomography
PET-CT was acquired using a PET-CT (GE Discovery VCT type, America) scanner. Subjects fasted for 4–6 h before imaging and their blood glucose was controlled to below 7 mmol/L. After a quiet rest for 15 min, they were injected intravenously with fluorodeoxyglucose (18F-FDG) in the arm at 7.4 MBq/kg. Then, 20–30 min after injection, 800–1000 ml of water was drunk. Patients urinated at 60 min postinjection and PET-CT was then performed. First, a spiral CT scan was acquired with a slice thickness of 3–4 mm, slice spacing of 3–4 mm, tube voltage of 120 kV, and tube current of 350 mA. The PET was then performed using a three-dimensional (3D) acquisition, with each bed collected for 3 min and a total of 6–8 beds being collected. After data acquisition, images were reconstructed using iterative reconstruction, the PET image was attenuation-corrected using the CT image, and the CT images and corrected PET were automatically fused to obtain axial, coronal, and sagittal PET-CT fusion images.
Image processing and data analysis
The overall workflow is summarized in [Figure 1].
|Figure 1: A flowchart for evaluation of the activity of patients with hepatic alveolar coccidiosis using the radiomics method|
Click here to view
Positron emission tomography–computed tomography observation items and evaluation criteria
A workstation was used to perform semi-quantitative analysis and automatically calculate the FDG standardized uptake values (SUV) of lesions. A previous study showed that an early maximum SUV of 4.88 ± 1.86 indicated that the marginal band of the lesion was a proliferative invasion zone and that it will further invade normal liver parenchyma and cause disease progression. Therefore, the lesion area was carefully assessed and an SUV higher than 5 was used to determine that a lesion was active, whereas a maximum SUV lower than 5 was taken to indicate an inactive lesion. Using the SUV values as the reference standard to distinguish between active and inactive lesions, 90 patients were diagnosed with active lesions and 46 with inactive lesions.
Different MRI scanners with different field strengths were used, and therefore, the images were preprocessed to obtain more robust radiomics features. The image preprocessing consisted of two steps:
- Step 1: To eliminate the inherent dependence of the radiological features on voxel size, the voxel size was normalized using a resampling method with linear interpolation
- Step 2: To minimize MRI intensity changes, the image intensity was normalized using the following formula (whereis the image original intensity; represents the normalized image intensity; is the mean image intensity; expresses the image intensity variance; is a variable parameter, set to 1 by default):
The HAE cases which met the criteria were included, and the Radcloud software (http://www.radcloud.cn/) of Huiyi Huiying's radiomics cloud platform was for image processing. The 5-mm slice thickness T2WI sequence was imported into the software platform in DICOM format and the window width was set to 1860 and the window position to 1070. Four imaging physicians manually outlined a region of interest (ROI) around the edge of the lesion in a slice-by-slice manner. Another senior radiologist with more than 10 years of experience in abdominal diagnosis checks the delineation results (if there was any objection, the senior imaging doctor determined the lesion boundary) and the computer automatically interpolated across the single slice ROIs to generate a 3D volume of interest (VOI) of the lesion, as shown in [Figure 2] and [Figure 3].
|Figure 2: Delineation and 3D modeling of the region of interest. (a) A 47-year-old male with an active hepatic alveolar echinococcosis lesion on T2WI. (b) Computer-generated 3D volume of the 47-year-old male with active hepatic alveolar echinococcosis lesion. 3D: Three-dimensional, T2WI: T2-weighted images|
Click here to view
|Figure 3: Delineation and 3D modeling of the region of interest. (a) A 39-year-old female with an inactive hepatic alveolar echinococcosis lesion on T2WI. (b) Computer-generated 3D volume of the 39-year-old female with inactive hepatic alveolar echinococcosis lesion. 3D: Three-dimensional, T2WI: T2-weighted images|
Click here to view
Feature extraction and screening
A total of 1409 high-throughput data features based on different feature and filter classes were extracted from each VOI. The features were divided into four groups. Group 1 (first-order statistics) consisted of 18 feature descriptors that quantitatively depict the distribution of voxel intensities within the MR image (such as peaks, means, and variances) through commonly used and basic metrics. Group 2 (features based on shape and size) contained 14 3D shape features reflecting the shape and size of the tumor area. Group 3 (second-order texture features) contained 75 texture features that quantified differences in regional heterogeneity, such as the gray level co-occurrence matrix, gray level run length matrix, and gray level size-zone matrix. Group 4 (high-order filter features) included 1302 filter class features extracted by filtering images with logs, index, and wavelet filters, and then extracting first-order statistics based on the filtered images and texture characteristics. Considering the active group and the active group data ratio imbalance, Synthetic Minority Oversampling Technology (SMOTE) was applied to secondary samples in the experiment.
Owing to differences between the characteristics of the different properties, the characteristic value was normalized to ensure the training model converged. At the same time, to avoid the risk of model overfitting caused by random division of the training set and test set, five-fold cross-validation was used in the experiment to divide the data set into training sets and test sets; in each round of training, 80% of the data were randomly selected and used as a training set, with the remaining 20% used as a test set. In each of the training sets, the least absolute shrinkage and selection operator (LASSO) regression algorithm was used to reduce the dimension of the above features, selecting the most meaningful ones, and the correlation coefficient of the corresponding features is calculated.
When the number of features in a certain fold was >10, the 10 features with the highest correlation coefficients were selected as the final feature subset.
Model establishment and verification
Three machine learning models were constructed based on the extracted optimal image features: logistic regression (LR), multilayer perceptron (MLP), and support vector machine (SVM) models. The patients were divided into a training set and test set (the data set was divided into five parts, four of which were used for training and one for testing, in turn). Using five-fold cross-validation, the classification prediction results were obtained, and a receiver operating characteristics curve was established (ROC) to obtain the model cutoff value and calculate the area under the curve (AUC), confidence interval of the AUC (95% CI-AUC), accuracy (ACC), sensitivity, and specificity. All evaluation values are the average of five test results, and were used to compare the classification results of the different classifiers to build a relatively good activity evaluation model for HAE.
Standardization, feature selection, and model building were performed using Python 3.6 (https://www.python.org/). This study used the “scikit-learn” (https://scikit-learn.org/) and “matplotlib” (https://matplotlib. Org/) packages. Statistical analysis of clinical information was performed using SPSS version 17.0 (IBM Corp., Armonk, NY, USA). Normality tests were performed on the associated measurement data, and if a normal distribution was met, the data were described as mean ± standard deviation and independent sample t-tests were used for comparisons. If a normal distribution was not met, data were expressed as the median (interquartile space) and were compared using a nonparametric test. Chi-square tests were used for comparisons of count data. The diagnostic performance of the machine learning classifiers was evaluated using ROC analysis (95% CI-AUC, specificity, and sensitivity).
| Results|| |
Basic patient information
The 136 patients had an average age of 39 ± 13 years and 72 were women (53%). According to the preoperative PET-CT examination results, 90 patients were assigned to the active group (47 women) and 46 patients to the inactive group (25 women). No obvious statistically significant differences in the clinical characteristics of sex, age, lesion location, and size were found between the active and inactive groups. The clinical data of the patients are shown in [Table 1].
Radiomics label establishment
A total of 1409 radiomics features were extracted from each T2WI VOI, and the data were divided into five different training sets for the five-fold cross-validation method (taking it in turns to use each part for testing). Dimensionality reduction analysis was performed on each training set using the LASSO method, with the LASSO method adopting ten-fold cross-validation and a maximum of 2000 iterations. To avoid overfitting of the model due to the high dimensionality of the characteristics, for groups with more than 10 features after screening, the 10 features with the largest absolute values were further filtered out through the ranking of the feature correlation coefficients, as shown in [Table 2].
|Table 2: Best features in each group in the five groups of feature screening experiments|
Click here to view
The sum of the best ten characteristic coefficients in the five groups of experiments of five-fold cross-validation constructs a bar graph, as shown in [Figure 4]. As can be seen from [Figure 4], the optimal features included first-order statistical features (n = 2) and texture features (n = 1) of the original image, as well as filter-transformed first-order statistical features (n = 16) and texture features (n = 29). The maximum feature operator (wavelet-HLH_firstorder_Maximum) in the wavelet-transformed first-order statistical features was particularly important, appearing in all five sets of experiments. With the largest cumulative correlation coefficient, it can be used as an effective image group marker for the assessment of HAE activity.
|Figure 4: Cumulative graph of the correlation coefficients of each group of optimal features in the five groups of feature screening experiments|
Click here to view
Radiomics difference analysis
A variety of machine learning algorithms were used to train the model and perform five-fold cross-validation. [Figure 5] and [Figure 6] show the ROC curves of three machine learning classifier models, and [Table 3] and [Table 4] summarize the diagnostic performance and model cutoff values for HAE activity in the five training sets and test sets. Overall, all three machine learning classifier models performed well, with the mean AUCs of the test set all being higher than 0.800. Of the three models, the MLP obtained the best discriminant effect for HAE activity prediction, with an average AUC on the training set of 0.925 ± 0.057, mean ACC of 0.866, mean sensitivity of 0.883, and mean specificity of 0.889. On the test set, the mean AUC, the mean ACC rate, the mean sensitivity, and the mean specificity were 0.830 ± 0.053, 0.817, 0.822, and 0.811, respectively.
|Figure 5: ROC curves of LR, MLP, and SVM machine learning classifiers in the training set (a) LR, (b) MLP, (c) SVM. ROC: Receiver operating characteristic, LR: logistic regression, MLP: Multilayer perceptron, SVM: Support vector machine|
Click here to view
|Figure 6: ROC curves for LR, MLP, and SVM machine learning classifiers in the test set. (a) LR, (b) MLP, (c) SVM. ROC: Receiver operating characteristic, LR: logistic regression, MLP: Multilayer perceptron, SVM: Support vector machine|
Click here to view
|Table 3: Logistic regression, multilayer perceptron, and support vector machine learning classifier model results in the training set|
Click here to view
|Table 4: Logistic regression, multilayer perceptron, and support vector machine learning classifier model results in the test set|
Click here to view
| Discussion|| |
Conventional radiomics-based tests applied to ultrasound, CT, and MRI can show the lesion range and calcification, but assessing the activity of the lesion by the naked eye alone can be difficult. This coupled with the heterogeneity caused by the coexistence of multiple components of HAE and the large individual differences of lesions makes it difficult to analyze and interpret images. Although it is the reference standard for evaluating the activity of HEA lesions, PET-CT has certain limitations preventing its wide application in clinical diagnosis and treatment. The radiation dose is slightly higher than that of ordinary examinations, and complicated prescanning preparations are required. Furthermore, accessing large-sized and expensive instruments such as PET-CT in remote areas is difficult, and areas with a high incidence of HAE are mostly remote pastoral areas, making its application scope very limited. The objective of this study was, therefore, to find a method to compensate for the above deficiencies and allow the evaluation of activity with no less ACC than PET-CT.
Liver bubble spherical lesions and the edge belt show uneven signals on conventional MRI, and most lesions show a mixed low signal on T1WI and T2WI; the low signal on T2WI is characteristic, with small vesicles showing a high signal on T2WI. MRI offers great advantages in soft-tissue contrast and submillimeter spatial resolution, allowing small vesicles and marginal bands to be shown more clearly, but it does not reflect changes in the microenvironment and cell metabolism of lesions associated with functional metabolism. MRI radiomics allowed the analysis of the internal environment of the lesion using conventional scans, and facilitated the construction of an HAE activity prediction model relying on imaging features.
Previous studies showed,, that quantitative analysis of MRI images combined with radiomics can better identify the characteristics of disease heterogeneity and reveal the key components of the focus phenotypes of multiple 3D lesions at multiple time points in the treatment process, and that it even has the potential for lesion classification, providing a further basis for clinical decision-making.,, In this study, by extracting and screening the radiomics features, we constructed an HAE activity prediction model that relied on imaging features. The optimal features included the intensity characteristics and texture characteristics of the original image, that is, the first-order statistical features and the second-order statistical features. The first-order statistical features describe the distribution of individual voxel values without considering spatial relationships and mainly include the maximum, median, minimum, and mean values of voxel intensities. The second-order statistical features measure the spatial arrangements of voxel intensities; that is, they measure the heterogeneity within lesions. Filter-transformed first-order statistical features and texture features, namely, higher-order statistical features extracted using a neighborhood grayness difference matrix, reflect the spatial relationships between three or more voxels and are considered to be very similar to the human image experience. Other studies showed that the most differentiated features are texture-based features. In this experiment, the maximum intensity characteristics in the ROI after wavelet filter transformation showed significant differences in the assessment of activity (wavelet-HLH_firstorder_Maximum, P < 0.001, independent sample t-test); the average 95% CIs were 0.142–0.260 in the inactive group and 0.117–0.172 in the active group, and the maximum gray intensity in the lesion area of the active group was smaller than that in the inactive group. Considering the imbalance of the data ratio between the active and inactive groups, secondary sampling was carried out using SMOTE. SMOTE is an enhanced sampling method, in which the new synthetic sampling calculation is based on the Euclidean distance of the variable. The synthetic case will have values similar to the existing case, rather than being just a simple copy of the oversampling, so that the representation of a minority class in the resulting dataset is added while reflecting the structure of the original case. Notably, however, the new synthetic data only appear in the training set. Finally, three different machine learning algorithms were selected to train the model. LR uses a logical function to estimate the probability of the relationship between the classification-dependent variable and the independent variable. SVM achieves a nonlinear mapping to the high-dimensional space through a kernel function, with the final decision function of the SVM only being determined by a small number of support vectors, and it is, therefore, suitable for solving small sample problems. MLPs are forward structures of artificial neural networks that can handle nonlinear separable problems and are widely used and strongly scalable. In the training process, a 50% cross-validation process was used to obtain more robust results. The results showed that the mean sensitivity of the training set and the test set was >0.75, the mean specificity was >0.8, and the mean AUC was >0.8, with the MLP achieving the optimal prediction performance. This may be because the deep neural network architecture of the MLP overcomes the limitation of locally optimal solutions, suggesting that the model has high efficiency and can construct the best predictive radiomics label, thereby providing a basis for evaluating HAE activity on MRI. The HAE activity prediction model relying on imaging characteristics not only is free from ionizing radiation but also has high economic efficiency and offers prediction performance highly consistent with that of PET-CT. We, therefore, believe that it will be an indispensable evaluation method in future HAE diagnosis and treatment.
| Conclusion|| |
This study has limitations. First, it is of a single-center retrospective design, and the number of inactive lesions is small. To improve the model prediction efficiency, it will be necessary to expand the sample size and perform further multicenter prospective randomized double-blind trials in the future. In summary, this study constructed an HAE activity prediction model relying on imaging characteristics that provides valuable information for the diagnosis and treatment of clinical hepatic coccidiosis and its prognostic prediction.
We thank Liwen Bianji (Edanz) (www.liwenbianji.cn) for editing the language of a draft of this manuscript.
Financial support and sponsorship
National Natural Science Foundation of China (81974263).
Conflicts of interest
There are no conflicts of interest.
| References|| |
Turkhan A, Wen H, Shao YM. Comprehensive evaluation of preoperative radical resection of hepatic alveolar Hydatid. J Hepatobiliary Pancreat Surg 2006;14:154-6.
Hydatid Surgery Specialty Committee of Surgeons Branch of Chinese Medical Doctor Association. Expert consensus on diagnosis and treatment of hepatic cystic and alveolar echinococcosis (2019 edition). Chin J Digest Dis Surg 2019;18:711-21.
Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al.
Radiomics: Extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6.
Han D, Li QH, Cai W, XiaYW, Ning J, Huang F, et al.
Research and application of artificial intelligence in medical imaging. Big Data Res 2019;5:39-67.
Liu M, Ma X, Shen F, Xia Y, Jia Y, Lu J. MRI M, Ma radiomics nomogram to predict synchronous liver metastasis in primary rectal cancer patients. Cancer Med 2020;9:5155-63.
Oikonomou EK, Siddique M, Antoniades C. Artificial intelligence in medical imaging: A radiomic guide to precision phenotyping of cardiovascular disease. Cardiovasc Res 2020;116:2040-54.
Nie P, Yang G, Wang Z, Yan L, Miao W, Hao D, et al.
A CT-based radiomics nomogram for differentiation of renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma. Eur Radiol 2020;30:1274-84.
Xiong Q, Zhou X, Liu Z, Lei C, Yang C, Yang M, et al.
Multiparametric MRI-based radiomics analysis for prediction of breast cancers insensitive to neoadjuvant chemotherapy. Clin Transl Oncol 2020;22:50-9.
Li XH, Zhang QZ, QinYD, Xie B,Amina I. 18F-FDG PET/CT to evaluate the therapeutic effect of liposomal albendazole on hepatic alveolar echinococcosis. Chin J Nucl Med Mol Imaging 2018;38:101-3.
Ye S, Zhao JQ, Jiang Y, Han Y, Xiao H, Li XY, et al
. Calcification characteristics and biological activity of hepatic alveolar echinococcosis. Chin J Rad 2018;52:283-5.
Liu WY, Jiang Y, Wang J, Zhang TL, Xiong XX, Li L, et al.
Imaging diagnostic criteria of hepatic echinococcosis in China. Radiol Infect Dis 2020;7:35-42.
Carré A, Klausner G, Edjlali M, Lerousseau M, Briend-Diop J, Sun R, et al.
Standardization of brain MR images across machines and protocols: Bridging the gap for MRI-based radiomics. Sci Rep 2020;10:12340.
de Leon AD, Kapur P, Pedrosa I. Radiomics in kidney cancer: MR imaging. Magn Reson Imaging Clin N Am 2019;27:1-13.
Li Z, Mao Y, Huang W, Li H, Zhu J, Li W, et al.
Texture-based classification of different single liver lesion based on SPAIR T2W MRI images. BMC Med Imaging 2017;17:42.
Zhao L, Gong J, Xi Y, Xu M, Li C, Kang X, et al.
MRI-based radiomics nomogram may predict the response to induction chemotherapy and survival in locally advanced nasopharyngeal carcinoma. Eur Radiol 2020;30:537-46.
Avanzo M, Wei L, Stancanello J, Vallières M, Rao A, Morin O, et al.
Machine and deep learning methods for radiomics. Med Phys 2020;47:E185-202.
Deng Y, Ming B, Zhou T, Wu JL, Chen Y, Liu P, et al.
Radiomics model based on MR images to discriminate pancreatic ductal adenocarcinoma and mass-forming chronic pancreatitis lesions. Front Oncol 2021;24:620981.
Alderson PO, Summers RM. The evolving status of radiomics. J Natl Cancer Inst 2020;112:869-70.
Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images are more than pictures, they are data. Radiology 2016;278:563-77.
Xie K, Sun HF, Lin T, Gao LG, Sui JF, Ni XY, et al.
Research progresses in feature extraction of radiomics. J Med Eng Technol 2017;33:1792-6.
Ehrhart IC, Parker PE, Weidner WJ, Dabney JM, Scott JB, Haddy FJ. Coronary vascular and myocardial responses to carotid body stimulation in the dog. Am J Physiol 1975;229:754-60.
Bagher-Ebadian H, Janic B, Liu C, Pantelic M, Hearshen D, Elshaikh M, et al.
Detection of dominant intra-prostatic lesions in patients with prostate cancer using an artificial neural network and MR multi-modal radiomics analysis. Front Oncol 2019;26:1313.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6]
[Table 1], [Table 2], [Table 3], [Table 4]