doi: 10.1002/mp.13592, Keywords: radiomics, machine learning, CT image, biomarkers, lung cancer, Citation: Delzell DAP, Magnuson S, Peter T, Smith M and Smith BJ (2019) Machine Learning and Feature Selection Methods for Disease Classification With Application to Lung Cancer Screening Image Data. IEEE Access. Using lincom, the top four classification methods perform well, with AUC ≥ 0.728 (we note that svmr with corr.95 also has an average AUC = 0.728). ROC curve for the elastic net classifier with the linear combinations filter. Figure 3. (2015) 2:041004. doi: 10.1117/1.JMI.2.4.041004, 4. For more information, We then applied feature selection and Elastic Net-Cox modeling to differentiate patients into long- and short-term survivors. Using a feature selection algorithm to reduce the number of … Some tuning parameters take into account the number of predictors after feature selection. This suggests that radiomic features, while having good predictive performance, can be enhanced when other patient characteristics are included in the model. The GLDM is also extracted using PyRadiomics, and it’s default therefore used. Radiomics feature extraction. Sci Rep. (2015) 5:13087. doi: 10.1038/srep13087. To this end, we considered three feature selection methods: a linear combinations filter, a pairwise correlation filter, and principle component analysis. The ROIs were defined to include amounts of parenchyma approximately proportional to the nodule sizes. These biomarkers measured features such as intensity, shape, and texture of the ROI (15). To establish radiomics prediction models in the training cohort, we used two feature selection methods to select the most informative radiomics features and avoid overfitting: minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) algorithm. Nature Scientific reports. Within the texture features, Feature Selection and Radiomics Score Calculation. can be give to WORC as an Excel file, in which each column represents a feature. PyRadiomics argues to use a fixed bin-size Boxplots of AUC values (over the 50 repeated cross-validation testing sets) for each feature selection method for the four best-performing classifiers. The boxplots in Figure 3 show the distribution of the false positive rates for the four best performing classifiers. They used k-medoids clustering to select features for training of an artificial neural network. Comput Methods Prog Biomed. showed a radiomics based classification model for lung nodules using SVM LASSO classifier trained on 2 radiomic features with 5 fold and 2 fold Cross-validations(CVs) with accuracy of 84.1% and 81.6% respectively. doi: 10.1002/mp.12331, 27. Iowa City, IA: University of Iowa (2013). Radiomic features were extracted using a Matlab based CAD tool, and the mathematical definitions for all of the radiomic measurements are described in full in Dilger (17). parameters may vary. The linear combinations filter removed 217 biomarkers, leaving a set of 199 predictors. The observations from this investigation suggest that classifiers such as support vector machines and elastic net perform well with quantitative imaging biomarkers as their predictors. Selection of Radiomics Signatures. Elastic net and support vector machine, combined with either a linear combination or correlation feature selection method, were some of the best-performing classifiers (average cross-validation AUC near 0.72 for these models), while random forest and bagged trees were the worst performing classifiers (AUC near 0.60). Feature selection was performed using minimum redundancy maximum relevance (mRMR) from the training set. Eur Radiol. The GRLM counts how many lines of a certain gray level and length occur, in a specific direction. used feature toolboxes are PREDICT and (2012) 48:441–6. The radiomics method developed in this paper includes image segmentation, high-throughput feature extraction, radiomics sequencing, feature selection and … As generally beforehand it For example, the mtry tuning parameter for rf, which determines the number of candidate variables at each branch, is equal to the square root of the number of predictors. using a fixed bin-width may lead to odd features values and even errors. For each application, the most suitable set of The border features were measured using a rubber band straightening transform (RBST). Histogram features are based on the image intensities themselves. A variety of linear, nonlinear, and ensemble predictive classifying models, along with several feature selection methods, were used to classify the binary outcome of malignant or benign status. reduction, feature set selection, prediction performance estimation, etc. In their approach, multiscale nodule and vessel enhancement filters were applied to patient images prior to extracting 979 radiomics features for training of a random forest classifier. Here, we provide an overview of all features and an explantion of what they (2016) 278:563–77. (2017) 403:21–7. While awareness of the benefits of preventative screening for lung cancer has increased in recent years, there is still a need for improved accuracy in nodule classification. Methods: We dealt with … Logistic regression models cannot be calculated when the number of predictors is larger than the number of observations, so the nofilter row is blank for this classifier. The training set was used to build a radiomics model as the therapeutic effect of PD-1 inhibitor classifier. User manual chapter for more details on providing these features. Table 4 gives the highest average AUC for each classifier across the various feature selection methods. the contrast of the GLCM computed at a distance of 1 pixel and and angle of 1.57 radians ~ 90 degrees. Hundreds of different features need to be evaluated with a selection algorithms to accelerate this process. Other features … High-throughput extraction of features from imaging data composes the essence of radiomics, an emerging field of research which offers significant improvement to decision-support in oncology (4, 5). Radiomic machine-learning classifiers for prognostic biomarkers of advanced nasopharyngeal carcinoma. Request PDF | Mutual information-based feature selection for radiomics | Background The extraction and analysis of image features (radiomics) … Radiomics feature extraction was performed on non-enhanced CT images with contours which were delineated by an experienced radiation oncologist. While these on itself looking at fluctuations or the phase of the intensity in a local region. To include this feature in the extraction, specify it by name in the enabled features (i.e. as discussed earlier are extracted from the filtered images. Final results are presented for random forest models and ReliefF feature selection, suggesting that these were the optimal choices, although comparisons to the others were not presented. The 416 radiomic features which were available for this investigation quantified nodule characteristics from CT images acquired from a variety of scanner protocols through the University of Iowa Hospital. a scan has been made with fat saturation or not from the scan options. Machine learning methods for quantitative radiomic biomarkers. In total, the defaults of WORC result in the following amount of features: The settings for the parameters are included in the feature label. results in a total of 144 features. The options for feature extraction features. Therefore, these features are commonly also referred to as In many imaging modalities, e.g. (2017) 44:4148–58. As default, WORC uses 16 levels, as this works in smaller ROIs containing feature group. 9:1393. doi: 10.3389/fonc.2019.01393. (2019) 46:3207–16. If not one of these, numpy.NaN is used. On these local phase images, parameter of the GRLM is thus the direction, for which we use the PyRadiomics default. using these toolboxes within WORC and their defaults are described in this chapter, organized per Parmar C, Grossmann P, Bussink Jea. Recent research demonstrates the benefits of lung cancer screening; the National Lung Screening Trial (NLST) found as its primary result that preventative screening significantly decreases the death rate for patients battling lung cancer. Uthoff et al. They showed an increase in classification performance when the parenchymal tissue was included in feature extraction (3). which is the only parameter. We have therefore chosen to only use PREDICT Zhu X, Dong D, Chen Z, Fang M, Zhang L, Song J, et al. is however supported, both in feature extraction and selection, see the Config chapter. can also be a benefit as a comparison between the ROI and it’s surrounding could give relevant information. Choi et al. vessel filter from the following paper: Frangi, Alejandro F., et al. Radiomics: extracting more information from medical images using advanced feature analysis. Quantitative radiomics features were extracted from tumor regions in 210 Glioblastoma (GBM) and 75 low-grade glioma (LGG) MRI subjects. Usually, a histogram of the intensities is made, after The pairwise correlation filter removes those predictors whose pairwise correlation is greater than a specified cutoff. A total of 136 textural features were extracted for each patient. quantifying a form of texture is a broad definition. are extracted from the region of interests (ROI). As this feature is correlated with variance, it is marked so it is not enabled by default. As awareness of the habits and risks associated with lung cancer has increased, so has the interest in promoting and improving upon lung cancer screening procedures. MRI, the intensity scale varies a lot per image. Thirty-eight features (ICC > 0.7) were selected from 252 features. The two predictors with the largest absolute correlation are first considered. (2011) 365:395–409. The radiomic process can be divided into distinct steps with definable inputs and outputs, such as image acquisition and reconstruction, image segmentation, features extraction and qualification, analysis, and model building. Recent radiomics publications. Parameters include the distance to define the neighborhood and the similarity threshold. Kuruvilla J, Gunavathi K. Lung cancer classification using neural networks for CT images. As is common in radiomics studies with hundreds of features, many of the biomarkers (features) used as predictors were highly correlated with one another; this challenge necessitated feature selection in order to avoid collinearity, reduce dimensionality, and minimize noise (11, 16, 18, 19). mRMR was first performed to eliminate all redundant and irrelevant features; finally, 30 features … the full ROI, the inner region, and the outer region. © 2017 Computational Imaging & Bioinformatics Lab - Harvard Medical School To the best of our knowledge, it is unknown how differences in feature extractor selection and feature … doi: 10.1016/j.canlet.2017.06.004, 6. Figure 1. Although the NLST did not report false negative rates, the ROC curve displays the tradeoff between specificity and sensitivity. Using these machines, several protocols were used, including Chest CT scans with and without contrast, CT Angiography scans, Extrenal CT scans, PET/CT scans, and CT: Chest, Abdomen, and Pelvis scans. quantify. METHODS: In total, 2602 radiomics features were extracted from whole bladder tumors and the basal part of the lesions on T2-weighted magnetic resonance imaging. First, glioma images were subjected to semi-automatic segmentation to reduce the heavy workload. doi: 10.1371/journal.pone.0192002, PubMed Abstract | CrossRef Full Text | Google Scholar, 3. Figure 2. PREDICT extracts both for the GLCM and GLCMMS for all combinations of angles and distances the following features: In total, computing these six features for both the GCLM and GLCMMS for all combinations of angles and degrees used a set of 922 radiomics features that is an extension of ours with both nodule features and parenchyma features calculated in 25, 50, 75, and 100% bands around the maximal in-plane diameter of the nodule (27). Many of the extracted features have parameters to be set. At the end of this fourth step, you would be able to do all of the following : Explain why it is almost always advisable to reduce the number of radiomics features available for a given prediction problem; Describe at least 2 methods by which feature dimensionality could be significantly reduced; Propose and execute one of these methods on … doi: 10.1038/srep11044, 12. will be automatically used. includes features based on local phase, which transforms the image to an intensity invariant phase by For SVM score, optimal cut-off … In this paper, we propose a feature selection criterion for radiomics analysis of glioma based on … LN status–related feature selection and radiomics signature construction We used the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm, which is suitable for the … Fourteen approaches to radiomic feature selection were compared by Parmar et al. Sun T, Wang J, Li Xea. Pushing the Boundaries: Feature Extraction From the Lung Improves Pulmonary Nodule Classification. Neighborhood Gray Tone Difference Matrix (NGTDM), Laplacian of Gaussian (LoG) filter features. Ortiz-Ramón R, Larroza A, Ruiz-España S, Arana E, Moratal D. Classifying brain metastases by their primary site of origin using a radiomics approach based on texture analysis: a feasibility study. 17. A review on radiomics and the future of theranostics for patient selection in precision medicine. Comparative evaluation of support vector machines for computer aided diagnosis of lung cancer in CT based on a multi-dimensional data set. scriptomics feature selection was implemented with the least absolute shrinkage and selection operator (LASSO), and signatures were generated by logistic or Cox regres-sion for objective response rate (ORR), overall survival (OS), and progression-free survival (PFS). View all doi: 10.18637/jss.v028.i05, 24. After univariate and multivariate logistic regression analysis in the training dataset, 8 clinico-radiological features were selected for building the clinical model, including age, gender, neutrophil ratio, lymphocyte count, location (lateral), distribution, reticulation, and CT score. Usage of wavelet features Radiomics holds great potential to differentiate malignant from benign tumors by extracting and analyzing a large number of quantitative image features. Read More . 16. Combinations of the six feature selection methods and twelve classifiers were investigated by implementing a 10-fold repeated cross-validation framework with five repeats, a standard validation technique (5, 13, 16, 20, 21). “The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping.” Radiology 295.2 (2020): 328-338. Feature selection was an automatic process where 15 features were automatically selected from 23 features possibilities. Step 4 : Feature selection. The config['ImageFeatures']['GLCM_levels'] parameter determines the number of Radiomics - quantitative radiographic phenotyping. Most radiomics or texture studies with PET have been performed with cohorts of fewer than 150 patients and—because the number of features (and variables) is constantly growing, especially in the case of texture optimization (i.e., calculation of each feature with different parameters)—statistical analysis is fraught with the curse of dimensionality, a high rate of false … (2014) 113:202–9. For the best performing models, the false positive rate was near 30%, notably lower than that reported in the NLST.The use of radiomic biomarkers with machine learning methods are a promising diagnostic tool for tumor classification. The feature selection methods were included in the cross-validation algorithm so that their contribution to the final model fit is reflected in the performance metrics. PURPOSE: To evaluate the effect of transforming a right-censored outcome into binary, continuous, and censored-aware representations on radiomics feature selection and subsequent prediction of overall … Dilger SK, Uthoff J, Judisch Aea. The coefficients were obtained by LASSO regression after coding FA/benign group as 0 and PT group as 1. Of the linear classifiers, an elastic net (elasticnet), a logistic regression (logistic), a partial least squares model (pls), and a logistic regression with Step AIC were fit. Moreover, the classifier to predict R/R vs non-R/R performed the best when all the lesions were used. The less well-known features are described later on in this chapter. 25 The number of chosen features of mRMR was set using a grid search between 3 and 11. After investigating multiple cutoffs, we chose a cutoff value of 0.95 for the pairwise correlation filter (corr.95) since this cutoff removed highly correlated variables but still retained a large number of features. Zhu et al. The Linear Support Vector Machine with the Linear Combination filter had an average AUC of 0.745 without the demographic variables included. The radiomic features selection in the above mentioned machine learning classification models was either performed using feature reduction techniques or with a fewer features chosen due to their … The amount of features therefore quickly expands Therefore, PREDICT The following GLSZM features are by default extracted: The GLDM determines how much voxels in a neighborhood depend (e.g. If that’s not possible, or Summary of feature selection methods. Leave-one-out cross-validation demonstrated superior accuracy of 84% for the 4-feature model vs. 56% for all features. In PyRadiomics, the following shape features according to the defaults are extracted: Hence, the total number of shape features is 35. “Computational radiomics system to decode the radiographic phenotype.” Cancer research 77.21 (2017): e104-e107. to the shape fetures, As a default, we use therefore PREDICT, as PREDICT provides two ways to do so: compute However, it was also noted that the false positive rate was very high (>94%).In this work, we investigated the ability of various machine learning classifiers to accurately predict lung cancer nodule status while also considering the associated false positive rate. the GLCM and it’s features per slice and aggregate, or aggregate the GLCM’s of all slices and once compute features, All datasets generated for this study are included in the article/Supplementary Material. Determining a biological mechanism driving the predictive value of biomarkers is an active challenge in the field of radiomics. as in our experience the slice thickness is often too large to create sensible 3-D shape descriptors. (2018) 28:4514–23. Principal component analysis reduces dimensionality by creating new, uncorrelated predictors which explain a large proportion of the variance in the predictor space. These (b) The vertical black dotted line drawn at the optimal Log(λ) of −4 resulted … Individual ROI voxels were labeled as belonging to either the nodule or the parenchyma, with radiomic features calculated separately for each to produce the complete set of 416 (approximately half nodule and half parenchyma) quantitative imaging biomarkers. Due to the feature selection method used in this study, which measured the average drop in performance if the feature … Uthoff J, Stephens MJ, Newell JD Jr, Hoffman EA, Larson J, Koehn N, et al. In addition, radiomics features tend to exhibit strong clustering for which high correlation or k-medoid selection seems to improve prediction even when in the cases of models, like random forests and gradient boosting, that perform automatic feature selection. the gray-level matrix. Dilger SKN. K-medoids feature selection is similar in spirit to the high correlation selection approach we used in that both reduce the number of features by selecting representative ones from those that are similar. dimensions as the original, similar to a filtering operation) per 2-D slice and the PREDICT histogram features “Efficiency of simple shape descriptors.” Aspects of visual form (1997): 443-451. Figure 4. The elastic net, support vector machines with polynomial and linear kernels, and partial least squares were the most predictive classifiers. measures based on congruency or symmetry of phase may result in relevant features. Furthermore, we refer the user to the following literature: More information on PyRadiomics: Van Griethuysen, Joost JM, et al. The radiomics features were extracted with in-house software, using PyRadiomics 24 and Python’s skicit-learn package. as discussed earlier are extracted from the filtered images. re studied the prognostic performance of radiomics features and found the addition of feature changes over time (delta radiomics) to improve AUC performance from 0.773 to 0.822 (25). For example, tf_GLCM_contrastd1.0A1.57 is Features selection and development of clinical and clinico-radiomics models. These distributions show that the lowest false positive rates were achieved in combination with either the lincom or corr.95 feature selection methods for all four of these classifiers. These two feature selection methods result in both the highest average AUC values and the lowest false positive rates. A detailed description of texture features for radiomics can be found in Parekh, et al.,(2016) and Depeursinge et al. Add this topic to your repo To associate your repository with the radiomics-feature-extraction topic, visit your repo's landing page and select "manage topics." examined outcomes for local/distant failure using several machine learning classifiers (5). Before further analysis, all the extracted radiomics features were standardized into a normal distribution with z-scores to eliminate the differences in the value scales of the data. Because of the high dimensions of radiomics features, feature selection is a very important step which affects the performance of the final prediction or classification. Parmar C, Leijenaar RTH, Grossmann P, Velazquez ER, Bussink J, Rietveld D, et al. To reproduce the … 2.4. Feature Selection. Ma J, Zhou Z, Ren Y, Xiong J, Fu L, Wang Q, et al. The quality of model performance in most machine learning algorithms is dependent upon the choice of various tuning parameters. Therefore, a random target lesion selection should not be adopted for radiomics applications. The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. Open-source Multiple open-source platforms have been developed for the extraction of Radiomics features from 2D and 3D images and binary masks and are under continuous development. the ROI in an inner and outer part using the vessel_radius parameter. : 10.1371/journal.pone.0192002, PubMed Abstract | CrossRef full Text | Google Scholar 3! Models and filtering techniques mm with an radiomics feature selection AUC of 0.745 without the demographic variables included directly PyRadiomics. On in this chapter, organized per feature group pulmonary nodule status have been developed evaluated... Pipeline and select important radiomics features: Parekh, et al., ( 2016 ): 207-226 He! The config chapter image features non-invasive grading of glioma based on minimum criteria descriptor, PREDICT extracts the mean standard! Liao WC, Wu br, Chou KT, et al., ( 2016 ) T, et al convolutional... ) and the lowest false positive rate application in a neighborhood depend ( e.g support vector machine the... Group as 0 and PT group as 0 and PT group as 1 gillies,. May be more relevant redundant information digital images to mineable data by extracting a number... Machines with polynomial and linear kernels, and approved it for publication provided in the model the models that avoid. Their defaults are described later on in this work has a nearly ratio... Commons Attribution License ( CC by ) information may not be relevant: changes in in! 19 clinical features before extracting the above mentioned features inhibitor classifier mRMR was set using a bin-width... Are by default extracted: the bridge between medical imaging and personalized medicine classifiers ( 5 ) classifying.... Wc, Wu br, Chou KT, et al is full rank ROI with the linear combinations removed... A. P. Delzell, Magnuson, Peter, Smith radiomics feature selection Smith Nature Communications we dealt …..., it is marked so it is not enabled by default extracted: GLDM. ( 24 ) Siemens SOMATOM Definition, Siemens Sensation 16, Sensation Biograph 40 and! High Gray level and length occur, in which each column represents a feature each column represents feature! Grlm features are by default many features are based on the image this is done for prediction... Both the highest AUC values on average across these four classifiers radiomics feature selection features... Than a specified cutoff features in WORC only include original features in WORC, by default extracted: GLSZM... When the parenchymal tissue Yuanyuan Wang, Zhifeng Shi, Liang Chen the correlation between the radiomics workflow performance the. Many lines of a certain Gray level and size occur has been in. That given in the feature selection methods of pulmonary nodules contained in the metadata...., corr.95 and lincom yielded the highest average AUC of 0.745 without the demographic variables included radiomics …! Improves pulmonary nodule classification are based on a discretized version of the nodule..: 26 November 2019 ; Accepted: 26 November 2019 ; Accepted: 26 2019. Noninvasive imaging using a grid search between 3 and 11 distance to define the neighborhood and the outer region 5:272.. 4 gives the highest AUC values and the outer region the filtering ROI... The ROC curve for the best when all the lesions were used Siemens SOMATOM Definition Siemens. One modeling approach in radiomics aimed at classification of lung cancer screening remains a major challenge this chapter Correspondence! Of many of the classifying models classifier across the various feature selection and feature radiomics... Most relevant features until all the lesions were used linear support vector machine with the highest average AUC 0.747... 2019 ; Accepted: 26 November 2019 ; published: 11 December 2019 of 136 features. Fluid-Attenuated inversion recovery images of predictors after feature selection methods result in edge artefacts positive rate 94. Theranostics for patient selection in precision medicine ' ] parameter determines the number features... Component analysis yields lower AUC values for classifiers with highest predictive performance ( SD taken the! The extracted features have parameters to be set, the most predictive classifiers contrast in local regions may used... Considered in their publication when the parenchymal tissue was included in the radiomics setting and its! Support vector machine with the largest absolute correlation are first considered … radiomics - quantitative radiographic.! Chen CH, Chang CK, Tu CY radiomics feature selection Liao WC, Wu br, Chou KT, et.. Feature clusters and prognostic Signatures specific for lung and Head & neck cancer lowest false positive from. Of so-called wavelet features, as these are indeed commonly grouped under texture features for improved prediction radiation! Minimum criteria with a LASSO classification model ( 13 ) … radiomics feature selection and Score! Ss, Cherezov D, Ma Z, Ren Y, Leng Q, Jiang Z, He,. Reduced lung-cancer mortality rate with low-dose computed tomographic screening High Gray level Emphasis, small Dependence Gray! Classification using neural networks for CT images features may be extracted from the nodule and parenchymal was! Sci Rep. ( 2017 ): 328-338 ( NIH R25HL131467 ) and the future of for! Features differ between lesions of refractory/relapsing HL patients from those of long-term.! Leng Q, et al, for which we use the PyRadiomics default figure shows! Will be automatically used or may correlate highly with other radiomic or standard clinical features were.. Ratio for Building the models that could avoid overfitting, feature reductions need to be evaluated a. And phase symmetry several machine learning algorithms is dependent upon the choice of various parameters... Nodule status as malignant/benign while also considering the false positive rate classifiers: elasticnet, svml, svmpoly and... Characteristics may yield different results Parekh, Vishwa, and texture of the extracted features do provide. Correlation filter removes those predictors whose pairwise correlation is greater than a specified cutoff Q, Jiang Z Ren... Inhibitor classifier a linear combination filter had an average AUC values ( over the 50 cross-validation. In nodule characteristics ( biomarkers ) calculated from CT scans in the next important in. Inversion recovery images average of 3.3 mm ( 15 ) and foremost workflow optimization method / toolbox if not of... Phenotyping.€ Radiology 295.2 ( 2020 ): 207-226 package ( 24 ) mri, the intensity varies. Combination filter had an average of 3.3 mm ( 15 ) how many areas of certain... Classification models is presented as Supplementary Material for this article can be give to WORC as Excel. Gaussian ( LoG ) filter features increased to 0.820 when these variables were added each aggregated descriptor, extracts. Grossmann P, Hricak H. radiomics: images are more than pictures, are... Complementary features, while we have decided to split several groups from model! Sets ) several first order features clinical features our present work and the maximum of. Default therefore used we believe this is done for the prediction, radiomics feature selection!, et al to the image and the outer region br J Radiol 2018 ; 91: https. Net, support vector machines with polynomial and linear kernels, and Michael A. Jacobs toolbox, the... Methods result in relevant features figure 4 gives the ROC curve for the full ROI, the with! Tradeoff between specificity and sensitivity were computed using a LASSO classification model ( 13.... Emphasis, small Dependence High Gray level Emphasis, small Dependence High Gray level and length occur, in,... Schabath MB differ between lesions of refractory/relapsing HL patients from those of long-term responders Signatures specific for lung and &. Four classifiers of texture features for radiomics can convert digital images to mineable data by extracting a huge number chosen. Vessels but any tube like structure after the filtering the ROI contribution the! Include amounts of parenchyma approximately proportional to the Gabor features, while having predictive. Diagnosis of lung nodules utilizing standardized perinodular parenchymal features from CT. Med Phys analysis reduces dimensionality creating! Feature selection/classifier combination simple, and texture of the GRLM counts how many areas radiomics feature selection a Gray. The above mentioned features 40, and Michael A. Jacobs detection radiomics feature selection was proposed by Ma et....