Access the full text.
Sign up today, get DeepDyve free for 14 days.
M. Rahim, Sung Kim, Hyeongryul So, Hyung Kim, G. Cheon, Eun Lee, K. Kang, D. Lee (2014)
Recent Trends in PET Image Interpretations Using Volumetric and Texture-based Quantification Methods in Nuclear OncologyNuclear Medicine and Molecular Imaging, 48
M. Desseroit, D. Visvikis, F. Tixier, M. Majdoub, R. Perdrisot, R. Guillevin, C. Rest, M. Hatt (2016)
Erratum to: Development of a nomogram combining clinical staging with 18F-FDG PET/CT image features in non-small-cell lung cancer stage I–IIIEuropean Journal of Nuclear Medicine and Molecular Imaging, 43
Paul Kinahan, J. Fletcher (2010)
Positron emission tomography-computed tomography standardized uptake values in clinical practice and assessing response to therapy.Seminars in ultrasound, CT, and MR, 31 6
J. Hartigan, M. Wong (1979)
A k-means clustering algorithm
M. Wilkerson, D. Hayes (2010)
ConsensusClusterPlus: a class discovery tool with confidence assessments and item trackingBioinformatics, 26
Song Chen, S. Harmon, T. Perk, Xuena Li, Meijie Chen, Yaming Li, R. Jeraj (2017)
Diagnostic classification of solitary pulmonary nodules using dual time 18F-FDG PET/CT image texture features in granuloma-endemic regionsScientific Reports, 7
D. Krstajic, L. Buturovic, D. Leahy, Simon Thomas (2014)
Cross-validation pitfalls when selecting and assessing regression and classification modelsJournal of Cheminformatics, 6
G. Cook, C. Yip, M. Siddique, V. Goh, S. Chicklore, A. Roy, P. Marsden, S. Ahmad, D. Landau (2013)
Are Pretreatment 18F-FDG PET Tumor Textural Features in Non–Small Cell Lung Cancer Associated with Response and Survival After Chemoradiotherapy?The Journal of Nuclear Medicine, 54
B. Efron (1979)
Bootstrap Methods: Another Look at the JackknifeAnnals of Statistics, 7
N. Hylton, L. Esserman, E. Pisano, Sandra Polin (2020)
American College of Radiology Imaging NetworkDefinitions
Wicher Bergsma (2013)
A bias-correction for Cramér’s and Tschuprow’sJournal of The Korean Statistical Society, 42
E. Segal, C. Sirlin, C. Ooi, A. Adler, J. Gollub, Xin Chen, Bryan Chan, G. Matcuk, Christopher Barry, Howard Chang, M. Kuo (2007)
Decoding global gene expression programs in liver cancer by noninvasive imagingNature Biotechnology, 25
P. Lambin, R. Leijenaar, T. Deist, J. Peerlings, E. Jong, J. Timmeren, S. Sanduleanu, R. Larue, A. Even, A. Jochems, Y. Wijk, H. Woodruff, J. Soest, T. Lustberg, E. Roelofs, W. Elmpt, A. Dekker, F. Mottaghy, J. Wildberger, S. Walsh (2017)
Radiomics: the bridge between medical imaging and personalized medicineNature Reviews Clinical Oncology, 14
J. Kekäläinen, I. Larma, M. Linden, J. Evans (2015)
Lectin staining and flow cytometry reveals female-induced sperm acrosome reaction and surface carbohydrate reorganizationScientific Reports, 5
T. Coroller, P. Grossmann, Ying Hou, E. Velazquez, R. Leijenaar, Gretchen Hermann, P. Lambin, Benjamin Haibe-Kains, R. Mak, H. Aerts (2015)
CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma.Radiotherapy and oncology : journal of the European Society for Therapeutic Radiology and Oncology, 114 3
Hsiang-Fu Yu, Fang-Lan Huang, Chih-Jen Lin (2011)
Dual coordinate descent methods for logistic regression and maximum entropy modelsMachine Learning, 85
L. Kaufman, P. Rousseeuw (1991)
Finding Groups in Data: An Introduction to Cluster Analysis
Aaron Defazio, F. Bach, Simon Lacoste-Julien (2014)
SAGA: A Fast Incremental Gradient Method With Support for Non-Strongly Convex Composite Objectives
Dong Liu, J. Nocedal (1989)
On the limited memory BFGS method for large scale optimizationMathematical Programming, 45
M. Desseroit, D. Visvikis, F. Tixier, M. Majdoub, R. Perdrisot, R. Guillevin, C. Rest, M. Hatt (2016)
Development of a nomogram combining clinical staging with 18F-FDG PET/CT image features in non-small-cell lung cancer stage I–IIIEuropean Journal of Nuclear Medicine and Molecular Imaging, 43
Rui Xu, S. Kido, K. Suga, Yasushi Hirano, Rie Tachibana, Keiichiro Muramatsu, Kazuki Chagawa, Satoshi Tanaka (2014)
Texture analysis on 18F-FDG PET/CT images to differentiate malignant and benign bone and soft-tissue lesionsAnnals of Nuclear Medicine, 28
C. Parmar, R. Leijenaar, P. Grossmann, E. Velazquez, J. Bussink, D. Rietveld, M. Rietbergen, Benjamin Haibe-Kains, P. Lambin, H. Aerts (2015)
Radiomic feature clusters and Prognostic Signatures specific for Lung and Head & Neck cancerScientific Reports, 5
(2013)
API design for machine learning software: experiences from the scikit-learn project
Virendra Kumar, Yuhua Gu, Satrajit Basu, A. Berglund, S. Eschrich, M. Schabath, K. Forster, H. Aerts, A. Dekker, D. Fenstermacher, Dmitry Goldgof, L. Hall, P. Lambin, Y. Balagurunathan, R. Gatenby, R. Gillies (2012)
Radiomics: the process and the challenges.Magnetic resonance imaging, 30 9
F. Murtagh, P. Legendre (2011)
Ward’s Hierarchical Agglomerative Clustering Method: Which Algorithms Implement Ward’s Criterion?Journal of Classification, 31
Joost Griethuysen, Andrey Fedorov, C. Parmar, A. Hosny, Nicole Aucoin, V. Narayan, R. Beets-Tan, J. Fillion-Robin, S. Pieper, H. Aerts (2017)
Computational Radiomics System to Decode the Radiographic Phenotype.Cancer research, 77 21
Paul Kinahan (2011)
PET/CT Standardized Uptake Values (SUVs) in Clinical Practice and Assessing Response to Therapy
S. Monti, P. Tamayo, J. Mesirov, T. Golub (2003)
Consensus Clustering: A Resampling-Based Method for Class Discovery and Visualization of Gene Expression Microarray DataMachine Learning, 52
C. Klifa, J. Carballido-Gamio, L. Wilmes, Anne Laprie, C. Lobo, Elizabeth Demicco, M. Watkins, John Shepherd, J. Gibbs, N. Hylton (2004)
Quantification of breast tissue index from MR data using fuzzy clusteringThe 26th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 1
Marjan Bolouri, S. Elias, D. Wisner, S. Behr, R. Hawkins, Sachiko Suzuki, Krysta Banfield, B. Joe, N. Hylton (2013)
Triple-negative and non-triple-negative invasive breast cancer: association between MR and fluorine 18 fluorodeoxyglucose PET imaging.Radiology, 269 2
Hui Li, Yitan Zhu, E. Burnside, Erich Huang, K. Drukker, K. Hoadley, C. Fan, S. Conzen, M. Zuley, J. Net, E. Sutton, G. Whitman, E. Morris, C. Perou, Yuan Ji, M. Giger (2016)
Quantitative MRI radiomics in the prediction of molecular classifications of breast cancer subtypes in the TCGA/TCIA data setNPJ Breast Cancer, 2
M. Nicolasjilwan, Ying Hu, Chunhua Yan, D. Meerzaman, C. Holder, D. Gutman, R. Jain, R. Colen, D. Rubin, P. Zinn, Scott Hwang, P. Raghavan, D. Hammoud, L. Scarpace, T. Mikkelsen, James Chen, O. Gevaert, K. Buetow, J. Freymann, J. Kirby, A. Flanders, M. Wintermark (2015)
Addition of MR imaging features and genetic biomarkers strengthens glioblastoma survival prediction in TCGA patients.Journal of neuroradiology. Journal de neuroradiologie, 42 4
Mark Schmidt, Nicolas Roux, F. Bach (2013)
Minimizing finite sums with the stochastic average gradientMathematical Programming, 162
C Klifa (2004)
Quantification of breast tissue index from MR data using fuzzy clusteringConf. Proc. Ieee. Eng. Med. Biol. Soc., 3
R. Mukhtar, C. Yau, M. Rosen, V. Tandon, N. Hylton, L. Esserman (2013)
Clinically Meaningful Tumor Reduction Rates Vary by Prechemotherapy MRI Phenotype and Tumor Subtype in the I-SPY 1 TRIAL (CALGB 150007/150012; ACRIN 6657)Annals of Surgical Oncology, 20
T. Win, K. Miles, S. Janes, B. Ganeshan, M. Shastry, R. Endozo, M. Meagher, R. Shortman, S. Wan, I. Kayani, P. Ell, A. Groves (2013)
Tumor Heterogeneity and Permeability as Measured on the CT Component of PET/CT Predict Survival in Patients with Non–Small Cell Lung CancerClinical Cancer Research, 19
E. Rakha, J. Reis-Filho, F. Baehner, D. Dabbs, T. Decker, V. Eusebi, S. Fox, S. Ichihara, J. Jacquemier, S. Lakhani, J. Palacios, A. Richardson, S. Schnitt, F. Schmitt, P. Tan, G. Tse, S. Badve, I. Ellis (2010)
Breast cancer prognostic classification in the molecular era: the role of histological gradeBreast Cancer Research : BCR, 12
H. Aerts, E. Velazquez, R. Leijenaar, C. Parmar, P. Grossmann, Sara Cavalho, J. Bussink, R. Monshouwer, Benjamin Haibe-Kains, D. Rietveld, F. Hoebers, M. Rietbergen, C. Leemans, A. Dekker, John Quackenbush, R. Gillies, P. Lambin (2014)
Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approachNature Communications, 5
Exploration of PET and MRI radiomic features for decoding breast Shih
M. Vaidya, K. Creach, Jennifer Frye, F. Dehdashti, J. Bradley, I. Naqa (2012)
Combined PET/CT image characteristics for radiotherapy tumor response in lung cancer.Radiotherapy and oncology : journal of the European Society for Therapeutic Radiology and Oncology, 102 2
P. Lambin, Emmanuel Rios-Velazquez, R. Leijenaar, S. Carvalho, R. Stiphout, P. Granton, C. Zegers, R. Gillies, R. Boellard, A. Dekker, H. Aerts (2012)
Radiomics: extracting more information from medical images using advanced feature analysis.European journal of cancer, 48 4
C. Parmar, P. Grossmann, J. Bussink, P. Lambin, H. Aerts (2015)
Machine Learning methods for Quantitative Radiomic BiomarkersScientific Reports, 5
Library for Large Linear Classification
Hui Li, Yitan Zhu, E. Burnside, K. Drukker, K. Hoadley, C. Fan, S. Conzen, G. Whitman, E. Sutton, J. Net, M. Ganott, Erich Huang, E. Morris, C. Perou, Yuan Ji, M. Giger (2016)
MR Imaging Radiomics Signatures for Predicting the Risk of Breast Cancer Recurrence as Given by Research Versions of MammaPrint, Oncotype DX, and PAM50 Gene Assays.Radiology, 281 2
M. Vallières, Carolyn Freeman, S. Skamene, I. Naqa, I. Naqa (2015)
A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremitiesPhysics in Medicine & Biology, 60
S. Partridge, S. Partridge, J. Gibbs, Ying Lu, L. Esserman, D. Tripathy, D. Wolverton, H. Rugo, E. Hwang, C. Ewing, N. Hylton (2005)
MRI measurements of breast tumor volume predict response to neoadjuvant chemotherapy and recurrence-free survival.AJR. American journal of roentgenology, 184 6
S. Partridge, Heumann Ej, N. Hylton (1999)
Semi-automated analysis for MRI of breast tumors.Studies in health technology and informatics, 62
www.nature.com/npjbcancer ARTICLE OPEN Exploration of PET and MRI radiomic features for decoding breast cancer phenotypes and prognosis 1 1 1 2 2 1 1 Shih-ying Huang , Benjamin L. Franc , Roy J. Harnish , Gengbo Liu , Debasis Mitra , Timothy P. Copeland , Vignesh A. Arasu , 3 1 1 1 1 1,4 1,5,6 John Kornak , Ella F. Jones , Spencer C. Behr , Nola M. Hylton , Elissa R. Price , Laura Esserman and Youngho Seo Radiomics is an emerging technology for imaging biomarker discovery and disease-specific personalized treatment management. This paper aims to determine the benefit of using multi-modality radiomics data from PET and MR images in the characterization breast cancer phenotype and prognosis. Eighty-four features were extracted from PET and MR images of 113 breast cancer patients. Unsupervised clustering based on PET and MRI radiomic features created three subgroups. These derived subgroups were −6 statistically significantly associated with tumor grade (p = 2.0 × 10 ), tumor overall stage (p = 0.037), breast cancer subtypes (p = 0.0085), and disease recurrence status (p = 0.0053). The PET-derived first-order statistics and gray level co-occurrence matrix (GLCM) textural features were discriminative of breast cancer tumor grade, which was confirmed by the results of L2-regularization logistic regression (with repeated nested cross-validation) with an estimated area under the receiver operating characteristic curve (AUC) of 0.76 (95% confidence interval (CI) = [0.62, 0.83]). The results of ElasticNet logistic regression indicated that PET and MR radiomics distinguished recurrence-free survival, with a mean AUC of 0.75 (95% CI = [0.62, 0.88]) and 0.68 (95% CI = [0.58, 0.81]) for 1 and 2 years, respectively. The MRI-derived GLCM inverse difference moment normalized (IDMN) and the PET-derived GLCM cluster prominence were among the key features in the predictive models for recurrence-free survival. In conclusion, radiomic features from PET and MR images could be helpful in deciphering breast cancer phenotypes and may have potential as imaging biomarkers for prediction of breast cancer recurrence-free survival. npj Breast Cancer (2018) 4:24 ; doi:10.1038/s41523-018-0078-2 INTRODUCTION features from PET and MRI with breast cancer phenotypes and prognosis. The association between the unsupervised clusters of In cancer management, multiple imaging modalities such as radiomic features and outcome data was evaluated using χ test of computed tomography (CT), magnetic resonance imaging (MRI), independence. The pairwise relationships between PET and MRI positron emission tomography (PET), and single photon emission radiomic features and breast cancer outcome were determined by computed tomography (SPECT) are often prescribed for tumor Spearman’s rank correlation coefficients (ρ) and proportion of detection, staging, and characterization. As a result, the collective variance explained by the predictor from multiple regression ( imaging data are information rich and can be extracted for in- r ) for ordered and unordered clinical outcome, respectively. In mreg depth analysis. Recent advances in radiomics have demonstrated addition, we also examined the predictive performance of the power of transforming imaging data into multi-dimensional radiomic features to recurrence-free survival (RFS) of up to 5 1,2 mineable radiologic features that are relatable to gene years following imaging and tumor grade. 3–5 expression pattern and have significant predictive/prognostic 3,6–8 power. However, determining the optimal use of multi- modality radiomic features to correlate with disease phenotypes, RESULTS molecular characteristics, and disease prognosis remains an open Study cohort problem. While radiomic features from anatomical images, such as This retrospective study included 113 patients diagnosed with CT, have shown significant potential in predicting survival breast cancer. The median patient age at diagnosis of primary outcome, and in associating with clinical and genomic features 2,3,9 tumor was 49 (range 21–96). Patient and tumor characteristics are of various cancers, there are few studies investigating radio- summarized in Table 1. mics derived from molecular imaging modalities such as PET/ 10–13 CT. There are even fewer studies of radiomics for the same Unsupervised tumor and feature clustering disease across imaging modalities such as PET and MRI. The added value of these multiple-order and multiple-dimension For consensus clustering based on PET and MRI radiomic features, image features remains largely unknown. In our study, we the number of clusters that consistently generated the largest carefully investigated the association of higher-order image change in the area under consensus cumulative distribution 1 2 Department of Radiology and Biomedical Imaging, University of California, San Francisco, CA, USA; School of Computing, Florida Institute of Technology, Melbourne, FL, USA; 3 4 Department of Epidemiology and Biostatistics, University of California, San Francisco, CA, USA; Department of Surgery, University of California, San Francisco, CA, USA; 5 6 Department of Radiation Oncology, University of California, San Francisco, CA, USA and Joint Graduate Group in Bioengineering, University of California, San Francisco and Berkeley, Berkeley, CA, USA Correspondence: Youngho Seo (youngho.seo@ucsf.edu) Received: 6 February 2018 Revised: 10 July 2018 Accepted: 18 July 2018 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. feature pattern of PET and MR images (P= 0.0085, χ test). Figure 3b, Table 1. A summary of patient demographic characteristics is shown c indicate that 76.6% of tumor cluster I were HR+/HER2+(Luminal B) and triple-negative tumors while 65.0% of tumor cluster III consisted Characteristics (N) Type No. of patients (%) of the HR+/HER2− (Luminal A) tumors and 25.0% of the HER2+ tumors were found in tumor cluster II. In addition, the tumor clusters Tumor Histology (N= 111) Ductual or lobular carcinoma in situ 5 (4.5) were statistically significantly associated with whether the disease Invasive ductal carcinoma (IDC) 98 (88.3) would recur, not recur, or was never disease free (P= 0.0053, χ test). Invasive lobular carcinoma (ILC) 5 (4.5) In Fig. 4c, 80% of the patients who were never disease free were Mixed IDC and ILC 3 (2.7) found in tumor cluster III. Tumor Grade (N = 104) 1 (well differentiated) 15 (14.4) Primary tumor stage (T-stage) and lymph-node stage (N-stage) 2 (moderately differentiated) 57 (54.8) did not reach statistical significance for their association with the 3 (poorly differentiated) 32 (30.8) radiomic features (p = 0.19, 0.14, respectively, χ test). In addition, T stage (N = 102) T0 32 (31.4) there was no evidence of association between the tumor clusters T1 33 (32.4) and tumor histology (p = 0.084, χ test). The association between T2 27 (26.5) the tumor clusters and the anatomical site of disease recurrence T3 10 (9.8) was not conclusive based on the data considered in this study (p N stage (N = 101) N0 62 (61.4) = 0.28, χ test). N1 32 (31.7) N2 4 (4.0) N3 3 (3.0) Pairwise relationship of radiomic features with breast cancer Overall stage (N= 104) 0 33 (31.7) outcome IA, IB, IIA 42 (40.4) Figure 5a indicates that the first-order statistics of PET image IIB 14 (13.5) entropy and PET-derived GLCM dissimilarity, entropy , and HIST GLCM IIIA, IIIB, IIIC 13 (12.5) difference average, and difference entropy were estimated to be IV 2 (1.9) positively correlated with tumor grade. The first-order statistics of Breast cancer subtype (N = 107) HR+ /HER2− 56 (52.3) PET image uniformity and PET-derived GLCM maximum prob- HR+ /HER2+ 15 (14.0) ability, energy , homogeneity, and inverse variance were GLCM HR-/HER2+ 15 (14.0) negatively correlated with tumor grade (|ρ|≈ 0.48). There was no HR-/HER2− 21 (19.6) correlation (ρ > 0.4) between the PET or MR radiomic features and Disease recurrence (N = 114) No recurrence 81 (71.1) T, N, or overall stage. Recur 23 (20.2) Figure 5b displays PET image texture features of difference Never disease free 10 (8.8) average, difference entropy, dissimilarity, sum average, and PET Recurrence site (N= 72) No recurrence 61 (84.7) SUV and SUV (r 0.10) contributed to the variance mean max mreg Local recurrence 1 (1.4) seen in the feature values among the breast cancer subtypes. For Distant recurrence 10 (14.9) recurrence-free survival, Fig. 5b indicates that the first-order Recurrence free in 1 year (N = 85) Recurrence free 75 (88.2) statistics of MR image mean and minimum and MR-derived GLCM Not Recurrence free 10 (11.8) average intensity, sum average, difference average, and dissim- Recurrence free in 2 years (N= 85) Recurrence free 68 (80.0) ilarity (r 0.10) contributed to the feature variance between mreg Not Recurrence free 17 (20.0) the patient groups who were and were not disease free within 2–5 Recurrence free in 3 years (N= 85) Recurrence free 67 (78.8) years. We also found that MR-derived GLCM IDMN, MR-derived Not Recurrence free 18 (21.2) GLCM IDN, and PET-derived GLCM cluster prominence (r ¼ mreg Recurrence free in 4 years (N= 85) Recurrence free 65 (76.5) 0.9–0.12) had contribution to the feature variance between the Not Recurrence free 20 (23.5) recurrence-free patient groups within 1 year. A summary of Recurrence free in 5 years (N= 85) Recurrence free 60 (70.6) Spearman’s rank correlation coefficients and proportion of Not Recurrence free 25 (29.4) variance from multiple regression were reported for all PET and For breast cancer subtype definition, HR+ denotes tumors with ER+ or PR MR image features and the clinical outcome in the supplemental Tables 1 and 2. Radiomics exploratory study with small sample size function (CDF) was 3. Table 2 gives a summary of χ -test of Based on 8 patients, supplemental Fig. 1 suggests that MR-derived independence statistics and cluster consensus for all breast cancer uniformity (ρ = 0.67) and tumor surface-to-volume ratio (ρ = outcomes. HIST 0.71) were positively correlated with Oncototype DX score while MR-derived entropy (ρ = −0.67) and GLCM autocorrelation (ρ HIST Association of radiomic features with breast cancer outcome = −0.64) were negatively correlated with Oncotype DX score. In The unsupervised clustering based on both PET and MR radiomic addition, supplemental Figs. 2 and 3 shows PET radiomics of the features in Fig. 1a shows that the tumor clusters were statistically primary tumor was consistent and associated with that of the −6 2 and significantly associated with tumor grade (p= 2.02 × 10 , χ - recurrent tumors for 6 out of 8 patients. test). Figure 1b indicates that 57.8% of tumor cluster I consisted of poorly-differentiated tumors (high tumor grade) while tumor Radiomic-based classification of recurrence-free survival (RFS) and clusters II and III were each associated with more differentiated tumor grade tumors (lower tumor grade). We observed a strong PET image Figure 6 shows a heatmap of the nested cross-validation feature pattern among tumor clusters for deciphering tumor grade. performance of several classification algorithms at predicting Tumor overall stage was statistically significantly associated with the RFS. The nested cross-validation shows that logistic regression tumor clusters (p= 0.037, χ test) in Fig. 2a. Figure 2bshows that with ElasticNet regularization and L1 regularization display the 50.0% of tumor cluster II were stage 2 tumors while 42.5% of tumor cluster I consisted of stage 0 tumors and 38.5% of tumor cluster III highest predictive performance with a mean AUC of 0.74 (95% CI were stage 3 tumors. Figure 3a shows that the breast cancer = [0.62, 0.88] and [0.61, 0.89], respectively) for predicting subtypes were statically significantly associated with the radiomic recurrence-free survival in 1 year. For ease of algorithm npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation 1234567890():,; Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. Table 2. A summary of χ test statistics (p-value and Cramer’s V), median cluster consensus (CC), and the optimal clustering algorithm is listed to describe the degree of association between the patient clusters with a given clinical feature Clinical variable Clustering algorithm # of samples p-value (χ test) Cramer’s V Median CC −6a Tumor grade HC, Spearman 104 2.02 × 10 0.39 0.72 Tumor histology PAM, Euc 111 0.084 0.22 0.94 T-stage HC, Spearman 102 0.19 0.21 0.77 N-stage KMdist, Spearman 101 0.14 0.22 0.73 Overall stage PAM, Pearson 104 0.037 0.28 0.83 Breast cancer subtype HC, Spearman 107 0.0085 0.28 0.77 Disease recurrence KMdist, Spearman 114 0.0053 0.25 0.73 Recurrence site PAM, Pearson 72 0.19 0.21 0.86 a 2 indicates there is statistical significance for the χ test of independence at the 5% level interpretability, we selected ElasticNet logistic regression in this dissimilarity and entropy and low texture energy ) could GLCM GLCM study for classifying RFS. The ElasticNet logistic regression has be predictive of poorly differentiated breast cancer. In addition, lower predictive performance at predicting recurrence free in 2 the PET and MR radiomics were found to be associated with breast years with a mean AUC of 0.68 (95% CI = [0.58, 0.81]). The cancer subtypes. In a study of 84 cases, Li et al., 2016 found that ElasticNet logistic regression using all PET and MR radiomics the enhancement texture from the first post-contrast MR images generated a mean AUC of 0.67 (95% CI = [0.58, 0.78]), 0.64 (95% CI were highly correlated to the molecular subtypes of breast cancer = [0.55, 0.75]), and 0.57 (95% CI = [0.47, 0.68]) at distinguishing (normal-like, luminal A and B, HER2-enriched, and basal-like). This patients being recurrence free in 3, 4, 5 years, respectively. In study suggests that PET and MR images with large texture predicting tumor grade, logistic regression with L2 regularization variation (large difference entropy and dissimilarity) along with PET SUV and MR peak enhancement could be predictive of and Lbfgs, Newtoncg, or Sag solver was found have the highest max breast cancer subtypes. The finding not only confirmed the result predictive performance with a mean AUC of 0.76 (95% CI = [0.72, 0.83]). in Li et al., 2016, but also added predictive potential of PET and Table 3 listed the PET and MR radiomic features that are MR radiomics over MR radiomics alone. Furthermore, breast dominant in predicting RFS and tumor grade using the optimal cancer consists of several tumor subtypes and MRI phenotypes logistic regression algorithm. The key radiomic features for including unicentric mass, multilobulated mass, area enhance- predicting RFS in 1 year are the MR-derived GLCM IDN, MR- ment with and without nodularity and septal spreading, which derived GLCM IDMN, and the PET-derived GLCM cluster promi- could explain the correspondence between large image texture nence. The radiomic features that were consistently dominant in variation and breast cancer subtypes. predicting RFS are the MR-derived GLCM sum average, MR- Our study also investigated the predictive performance of PET derived GLCM average intensity, MR minimum intensity, MR- and MR radiomics for breast cancer recurrence free status and derived GLCM IDN, and PET-derived GLCM cluster prominence. tumor grade. Instead of using 900+ radiomic features such as gray The key radiomic features for predicting tumor grade consisted of level size zone matrix features and wavelet-based features 3,14,18 mostly PET-derived GLCM features such as inverse variance and reported in previous studies, we extracted a limited number homogeneity along with PET-derived first-order statistics of PET of radiomic features from both PET and MR images, which SUV . provided a more succinct number of features (84) considering the mean limited sample size (N = 85) in this study. Even though we extracted the same type of radiomic features from both PET and DISCUSSION MR images, the multi-modality radiomic features were able to Higher-dimensional radiomic features were successfully extracted provide additional information since PET and MR images captured from both F-FDG PET and MR images among patients diagnosed different intrinsic information of tumor biology. Figure 5b shows with breast cancer. In this study, radiomics were clustered in an that MR-derived GLCM IDMN and IDN, and PET-derived GLCM unsupervised fashion; in other words, the clustering algorithm had cluster prominence were highly correlated with 1-year RFS. no prior knowledge of the tumor phenotypes and disease Similarly, MR-derived GLCM IDN and IDMN emerge as key features outcome. The unsupervised learning allowed exploration of any for predicting patient 1-year RFS (highest AUC from the ElasticNet potential relationship between the PET and MRI radiomics to logistic regression). In addition, MR mean and minimum intensity, breast cancer phenotypic behaviors and disease prognosis. We MR-derived GLCM average intensity, MR-derived GLCM sum found statistically significant association of the PET and MR average (r ¼ 0.09–0.10), and PET-derived GLCM cluster mreg radiomics clusters with breast cancer tumor grade, which was prominence (r ¼ 0.04–0.05), which were among the features mreg previously reported to have prognostic value for disease survival moderately correlated with RFS at 2–5 years, would likely play an 15 19 rate. Among those radiomic features positively associated with important role in RFS prediction. In a previous study, tumor size breast cancer tumor grade were the first-order statistics of PET and enhancement texture from DCE-MR images were effective at image entropy and SUV and the PET-derived GLCM features distinguishing the risk of breast cancer relapse and are also HIST var including dissimilarity, entropy , difference average, different confirmed in this study. In addition, this study shows that PET- GLCM entropy, and cluster prominence and tendency. Among those derived GLCM features such as inverse variance and homogeneity radiomic features negatively associated with breast cancer tumor were the key predictors of tumor grade, confirmed by the grade were the first-order statistics of PET image uniformity and univariate analysis (|ρ| = 0.48) and the nested cross validation. PET-derived GLCM maximum probability, energy , homogene- These PET-derived GLCM features were ranked above the first- GLCM ity, and inverse variance (|ρ| ≥ 0.45). This finding suggests that F- order PET image statistics such as PET SUV from nested cross mean FDG PET images large in asymmetry (high cluster prominence and validation of tumor grade classification. Therefore, a combination 18 st tendency), large in F-FDG uptake texture variation (high of PET and MR radiomics (both 1 -order statistics and GLCM Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24 Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. (a) (b) (c) Fig. 1 PET and MR radiomics vs. tumor grade heatmap. a A heatmap of the PET and MR radiomic features is shown with the corresponding tumor grade and the tumor clusters resulted from the optimized consensus clustering. Each column represents a tumor and each row represents a radiomic feature. The PET and MR radiomic features are shown as z-scores. b The proportion of different grade tumors is shown for each tumor cluster. The frequency is shown with respect to the total number of tumors in each tumor cluster category. c The proportion of different tumor clusters is shown for each tumor grade category. The frequency is shown with respect to the total number of tumors in each tumor grade category npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. (a) (b) (c) Fig. 2 PET and MR radiomics vs. tumor overall stage heatmap. a A heatmap of the PET and MR radiomic features is shown with the corresponding tumor overall stage and the tumor clusters resulted from the optimized consensus clustering. b The proportion of different tumor overall stages is shown for each tumor cluster category. The frequency is shown with respect to the total number of tumors in each tumor cluster category. c The proportion of different tumor clusters is shown for each tumor overall stage category. The frequency is shown with respect to the total number of tumors in each tumor overall stage category Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24 Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. (a) (b) (c) Fig. 3 PET and MR radiomics vs. breast cancer subtype heatmap. a A heatmap of the PET and MR radiomic features is shown with the corresponding breast cancer subtype and the tumor clusters resulted from the optimized consensus clustering. b The proportion of breast cancer subtypes is shown for each tumor cluster. The frequency is shown with respect to the total number of tumors in each tumor cluster category. c The proportion of different tumor clusters is shown for each breast cancer subtype. The frequency is shown with respect to the total number of tumors in each breast cancer subtype category npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. (a) (b) (c) Fig. 4 PET and MR radiomics vs. disease recurrence status heatmap. a A heatmap of the PET and MR radiomic features is shown with the corresponding disease recurrence status and the tumor clusters resulted from the optimized consensus clustering. b The proportion of different disease recurrence categories is shown for each tumor cluster. The frequency is shown with respect to the total number of tumors in each tumor cluster category. c The proportion of different tumor clusters is shown for each disease recurrence category. The frequency is shown with respect to the total number of tumors in each disease recurrence category Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24 Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. (a) (b) Fig. 5 Pairwise relationship of radiomics with breast cancer outcome. a A heatmap of Spearman’s rank correlation coefficients (ρ) between the PET and MR radiomic features and the ordered clinical outcome is shown. Only the radiomic features with |ρ| > 0.2 are displayed. b A heatmap of proportion of variance from multiple regression (r ) between the PET and MR radiomic features and the unordered clinical outcome is mreg illustrated. Only the radiomic features with r > 0.04 are shown mreg npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. Fig. 6 Heatmap of the predictive performance of radiomics to breast cancer outcome. A heatmap depicts the classification performance in AUC and 95% confidence interval for several classification algorithms at predicting recurrence-free duration of 1–5 years and tumor grade. SVM denotes support vector machine. The classification name for logistic regression is defined as [Reg][Solver]LogReg, where [Reg] specifies the regularization scheme and [Solver] is the solver algorithm. For example, L1LiblinearLogReg denotes logistic regression with L1- regularization using Liblinear solver features) could be more useful as prognosticator of breast cancer. clustering analysis. In particular, we determined that breast cancer Furthermore, feature selection for predictive performance may be tumor grade and breast cancer subtypes can be well characterized by the PET-derived GLCM features and 1st-order statistics. We more effective in our study due to the cross-validation process we st used rather than depending heavily on the correlation coefficients found that and 1 -order image statistics and image texture from the pairwise univariate analysis. features of the first post-injection DCE-MR image and PET images There are limitations to this study. Some factors may affect the have high potential for predicting recurrence-free survival of different outcome between the PET and MRI radiomics, including breast cancer and tumor grade. Findings from data exploration the fact that PET and MR images capture intrinsically different and initial predictive performance evaluation provide optimism for eventual construction of an effective predictive model based on biological and physiological mechanisms. The purpose of the study was to determine, not to compare, the predictive power of both PET and MRI radiomics for improved personalized disease the PET and MRI radiomics. Furthermore, the PET and MR images management and treatment planning. were resampled to the same isotropic voxel size for consistent image analysis. However, the image voxel upsampling likely METHODS introduced image interpolation effects, which may affect the Image datasets accuracy of radiomic features in measuring image information. In This study was a retrospective study of medical records and medical addition, the cross-validation was conducted with different images and qualified as exempt by the UCSF Institutional Review Board. machine learning algorithms for the initial predictive performance. We identified all patients who were diagnosed with invasive breast cancer The dataset used for this paper was limited by size for a study of between January 1st, 2005 and December 31st, 2009 and underwent both this scope. For future studies, we plan to obtain an independent breast dynamic contrast-enhanced (DCE) MR imaging and whole-body F- image dataset to validate our current findings and thereby further Fluorodeoxyglucose ( F-FDG) PET acquired as PET-CT examinations at evaluate the value of image radiomics in predicting disease different time at UCSF. All imaging studies were acquired prior to prognosis. We hope to expand the dataset used in Supplement treatment, including surgery, radiation, and/or chemotherapy. In addition Fig. 1 to investigate the role of PET and MR radiomics in predicting to images of primary tumors, PET images of patients diagnosed with breast cancer specific genomics. The difference in PET radiomics recurrent metastases (N = 8) were obtained to explore the difference in radiomics between the primary and recurrent tumors. The PET images between the primary and recurrent tumors (patient # 25 and 116 were acquired at more than 5 years after the diagnosis of primary disease. in Supplemental Figs. 2 and 3) will be further investigated with MR imaging was performed as previously described using either a 1.5- larger dataset as a key predictor for the course of treatment for Tesla (T) imaging system (Signa, GE Medical Systems, Milwaukee, WI) or a recurrent disease. 3-T imaging system (MagnetomVerio, Siemens Medical Systems, Erlangen, In summary, we investigated the benefit of PET and MRI Germany) with the patient in prone position. The DCE-MRI series consisted radiomics in deciphering breast cancer phenotypes and disease of a three-dimensional (3D), fat-suppressed, T1-weighted gradient echo prognosis. As an initial explorative investigation, this study sequence in accordance with the ACRIN 6657 imaging protocol. MR demonstrated the potential value of PET and MR image-derived imaging was acquired at three time-points: pre-contrast-injection, early radiomics in characterizing tumor phenotypes using unsupervised post-contrast-injection, and late post-contrast-injection. F-FDG PET/CT Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24 Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. images were performed with an integrated PET/CT system (Biograph 16, Table 3. The feature importance of the repeated nested cross- Siemens Medical Systems or Discovery VCT, GE Medical Systems). The PET/ validation with optimal logistic regression algorithm with PET and MR CT and MR images were reconstructed using the scanner-specific radiomic features set is summarized workstation. Outcome Important features Image segmentation, standardization, and pre-processing Disease free in 1 year (ElasticNet) MR GLCM IDN (99.1%) Tumor regions on MR images were identified using an established MR GLCM IDMN (84.1%) enhancement criteria of 70% applied to the first post-contrast image. PET GLCM cluster prominence (83.0%) This empirical threshold was based on visual agreement with radiological MR entropy (81.5%) 23 HIST assessments in clinical practice. Normal-appearing stromal tissue MRI mean intensity (77.5%) surrounding the tumor was subsequently defined as fibroglandular tissue MR GLCM sum entropy (76.2%) and was segmented from adipose tissue using a fuzzy C-means clustering MR GLCM sum average (74.7%) method. Tumors in the PET images were segmented semi-automatically MR GLCM average intensity (74.7%) using a region-growing algorithm (MeVisLab©, MeVis Medical Solutions MR minimum intensity (73.9%) AG). The segmented tumor regions were confirmed by trained radiologists MR GLCM difference entropy (72.0%) (S.B., M.D.). The in-plane image resolution ranged from 0.5 mm to 1.2 mm Disease free in 2 years (ElasticNet) MR mean intensity (98.2%) and 4.1 mm to 5.5 mm for MR and PET images, respectively. The axial MR GLCM sum average (98.1%) image resolution ranged from 0.5 mm to 2.8 mm and 2.0 mm to 5.6 mm for MR and PET images, respectively. For appropriate image feature MR GLCM average intensity (98.1%) comparison, all MR and PET images were resampled to the same voxel MR minimum intensity (96.6%) 3 3 dimension of 0.5 × 0.5 × 0.5 mm and 2.0 × 2.0 × 2.0 mm , respectively. PET MR maximum intensity (89.4%) images were converted into the unit of standard uptake value (SUV), MR GLCM IDN (87.5%) normalized by patient body weight and the decay-corrected injected MR GLCM difference average (87.1%) activity. MR GLCM dissimilarity (87.1%) PET SUV (86.3%) min MR tumor compactness2 (84.3%) Radiomic features Disease free in 3 years (ElasticNet) MRI mean intensity (98.9%) We defined 42 radiomic image features to characterize tumors in the MR GLCM sum average (98.4%) following categories: intensity (9), shape (8), and texture features (25). MR GLCM average intensity (98.4%) Table 4 shows the summary describing the radiomic features extracted in MR minimum intensity (96.8%) this study. Mathematical definitions of all radiomic features were described MR GLCM difference average (85.0%) in this previous study. For this explorative study, we extracted only GLCM MR GLCM dissimilarity (85.0%) texture features since they have been shown effective as a potential 26,27 MR maximum intensity (84.8%) imaging biomarker. The intensity features described the first-order statistics of the image signal intensity and histogram-based statistics, MR tumor compactness2 (83.6%) which characterize the distribution of the tumor intensity. The intensity PET tumor compactness2 (83.2%) histogram of the tumor region was generated with a fixed bin width of PET SUV (81.7%) min voxel intensity for all images. The shape features captured the three- Disease free in 4 years (ElasticNet) MR minimum intensity (94.3%) dimensional (3D) geometric attributes of the tumor. The texture features MR mean intensity (93.2%) provided spatial relationship between neighboring voxels within the MR GLCM sum average (91.3%) tumor region to quantify intra-tumor heterogeneity. The texture features MR GLCM average intensity (91.3%) were derived from gray level co-occurrence matrix (GLCM), which presents PET GLCM cluster prominence (85.6%) how combinations of discretized gray levels of neighboring voxels are MR GLCM IMC2 (85.5%) distributed along a given image direction. In this study, image features PET tumor compactness2 (82.6%) were extraction from MR images acquired at the first post-injection time MR maximum intensity (79.5%) point. The first-order statistics and GLCMs were generated from the PET MR tumor compactness2 (79.2%) and MR images discretized with a fixed voxel-intensity bin width of 0.1 and MR GLCM IDN (77.9%) 5.0 for PET and MR images, respectively. Generally, there are 26 connected Disease free in 5 years (ElasticNet) MR minimum intensity (92.0%) neighborhoods in 3D for GLCM, which yields 13 unique directions within PET GLCM cluster prominence (79.8%) the neighborhood for a voxel distance of 1. Thus, 13 GLCMs were PET GLCM IDN (78.7%) generated for each 3D image dataset, and the mean of the texture features MR GLCM IMC2 (78.4%) computed from the 13 GLCMs were reported for each tumor region. All PET tumor maximum 3D diameter (77.1%) image features were computed using in-house software based on Python MR mean intensity (74.6%) (version 2.7.14) and Insight Segmentation and Registration Toolkit (ITK, version 4.10.1). The value of radiomic features were validated with those MR GLCM sum average (70.2%) computed with Pyradiomics open-source software. MR GLCM average intensity (70.2%) MR GLCM IDN (69.8%) MR energy (69.0%) HIST Clinical dataset Binary Tumor Grade (L2LbfgsLogReg) PET GLCM inverse variance (90.6%) The following clinical data was collected from patient charts contained in PET GLCM homogeneity1 (85.6%) the electronic health system: tumor histologic type, tumor grade, estrogen PET GLCM homogeneity2 (83.7%) receptor (ER), progesterone receptor (PR), and human epidermal growth PET Entropy (79.5%) HIST factor receptor 2 (HER2) status. The breast cancer subtypes were then PET GLCM sum average (78.4%) grouped into the following categories where, additionally, hormone PET GLCM average intensity (78.4%) receptor (HR) status was defined as positive (+) when the ER or PR or PET SUV (78.2%) mean both receptors were positive on immunohistochemistry: HR+/HER2−,HR PET GLCM entropy (76.5%) +/HER2+, HR-/HER2+, HR-/HER2−. The primary tumor staging (T-stage), PET GLCM sum entropy (72.4%) regional lymph node staging (N-stage), and overall staging, as defined by the American Joint Committee on Cancer, as well as presence, site, and PET GLCM difference average (70.3%) date of disease recurrence and recurrence site were extracted from the The number in () is the proportion of the number of times that the feature institution’s cancer registry. The cancer recurrence status was categorized was considered ‘important’ during the repeated nested CV out of the as no recurrence, recurrence, never disease free. The recurrence site had maximum number of CVs (3000) the categories of no recurrence, any local recurrence, any distant recurrence, such as recurrence in bone or systemically. To investigate npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. Table 4. A summary describing the radiomic features extracted from the PET and MR images are shown Feature type Feature name Description First-order statistics Min, max Minimum and maximum of the image intensity values (FOstats) Mean, variance Skewness Measure of lopsidedness of the intensity distribution Kurtosis Measure of the heaviness of the tail of the intensity distribution Entropy Measure of randomness in an image HIST Energy HIST Uniformity Degree of image intensity having similar probability HIST Shape and size (SS) Volume Compactness1 and Compactness2 As a function of volume and surface area Maximum 3D diameter The largest pairwise Euclidean distance between voxels on the tumor surface Spherical disproportion Degree of similarity in surface area between the shape and that with a radius of a sphere with the same volume as the tumor Sphericity Surface area Surface-to-volume ratio Texture (TX) Autocorrelation Measure of texture fineness and coarseness Cluster prominence Measure of image asymmetry of the GLCM Cluster shade Measure of the skewness of the GLCM Cluster tendency Measure of voxel clusters of similar gray-level values Contrast Measure of the local variations presented in the image Correlation Measure of the linear dependency of image intensity of the neighboring voxels Difference entropy Measure of the variability in neighboring intensity value differences Difference average Relationships between voxel clusters with similar intensity values and voxel clusters with different intensity values Difference variance Measure of heterogeneity Average intensity The mean gray level intensity of the GLCM vertical or horizontal distribution dissimilarity Energy Measure of homogeneity of an image GLCM Entropy Measure of image texture randomness GLCM Homogeneity1 and Homogeneity2 Inverse difference moment normalized (IDMN) and Measure of the local homogeneity of an image inverse difference normalized (IDN) Inverse variance Maximum probability The number of most occurred pair of neighboring intensity values Sum average Average value of the GLCM Sum entropy Measure of randomness of the GLCM Sum variance High weight on the elements different from the GLCM average value Sum squares Measure of the neighboring intensity level pairs about the mean GLCM intensity level IMC1 and IMC2 the effectiveness of PET and MR radiomic features to predict the duration Data analysis until disease recurrence, the recurrence-free survival (RFS) was repeatedly For data exploration, we performed unsupervised clustering of tumors, dichotomized using cutoff times of 1, 2, 3, 4, and 5 years. The patients who using consensus clustering based on PET and MR radiomic features. were recurrence-free beyond the cutoff time were labeled 1, whereas Consensus clustering is a method that provides consensus across multiple those who were not recurrence-free were labeled 0. Furthermore, we runs of a clustering algorithm by subsampling data as a way to evaluate evaluated the value of PET and MR radiomic features to predict tumor the cluster stability and the best number of clusters for a given dataset. For a cluster class, a cluster’s consensus was computed as the average grade. The tumor grade was dichotomized such that those with tumor proportion of clustering runs in which two items are clustered together grade (T1) and (T2) were labeled 0 and those with tumor grade 3 (T3) and 4 between all pairs of items belonging to the same cluster. To determine (T4) were labeled 1. In addition, we obtained Oncotype DX score for 8 the optimal clustering algorithm, we performed consensus clustering with patients out of this study cohort to explore the pairwise relationship the following algorithms: hierarchical clustering with agglomerative ward between tumor genomic data and radiomics. All data analysis was linkage (HC), K-means (KM) on a data matrix, K-means on a distance performed on clinical data extracted from our clinical imaging database, 32 33 matrix (KMdist), and partitioning around medoids (PAM). We used 1- and there was no clinical trial associated with this study cohort. Pearson correlation (Pearson), 1 - Spearman correlation (Spearman), and 1- Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24 Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. Euclidean distance (Euc) as the dissimilarity measure. We performed the available due to them containing information that could compromise consensus clustering with resampling (10,000 iterations). The number of research participant privacy. Please contact L. Esserman for the ONCOTYPE clusters was estimated by the cluster number that gave the largest change DX score of the limited number of patients. The radiomics data extracted in area under the consensus cumulative distribution function (CDF). The from the PET and MR images along with the corresponding clinical median of the cluster’s consensus (median cluster consensus) was outcome in this study are available in this file (https://ucsf.box.com/s/ computed among all cluster classes for the optimal clustering setting dqopi5rgxc9u79zbjo53t6wai8dmf5uu). Each unique tumor is identified by (algorithms and the number of clusters). We performed the χ -test of the column name ‘ptid_side’. independence between the tumor cluster labels and each clinical feature for inference of data association. Cramer’sV were computed to measure the strength of association for the χ -test of independence. For each ACKNOWLEDGEMENTS clinical feature, the optimal clustering algorithm was selected as the one The study was supported in part by Department of Defense Grant W81XWH-17-1- that estimated the highest Cramer’s V between the tumor clusters and the 0033, Precision Imaging of Cancer and Therapy Program (PICT) in Departments of clinical feature. We used a significance level of 0.05 for detecting a Radiation Oncology, and Radiology and Biomedical Imaging, UCSF, and National statistically significant association in the χ -tests of independence. To Cancer Institute Grant R01 CA154561. facilitate the selection of radiomic features important to predict a clinical outcome, Spearman’s rank correlation coefficients (ρ) were computed to evaluate the strength and direction of association between an ordered AUTHOR CONTRIBUTIONS clinical outcome (tumor grade, stages, and Oncotype DX score) and a S.H., B.L.F., and Y.S. designed the study. N.M.H. and E.F.J. provided the breast MR radiomic feature. For an unordered clinical outcome, such as breast cancer image data and clinical and MR-related insights for breast cancer diagnosis and subtype, we fitted multiple regression models and used the proportion of prognosis. E.R.P. and L.E. provided the ONCOTYPE DX score for the limited number of variance explained by the predictor (r ) to indicate the strength of mreg patients in this study cohort. R.H. performed the PET tumor segmentation, managed association. Consensus clustering was performed using ConsensusCluster- 35 2 PET and MR images, and developed image processing software for this study. S.H. Plus implemented in R. The χ -test was performed using chi2_contigency performed all the data analysis, developed in-house software for extracting radiomics implemented in the Python Scipy statistics package. The multiple and data analysis, and writing of the manuscript. T.P.C. and V.A.A. extracted necessary regression and Spearman’s rank-order correlation coefficient were clinical data from the medical record and UCSF cancer registry. S.B. provided clinical implemented in R (version 3.3.2). insight for tumors extracted from the PET images. J.K. provided statistical consultation for all the analysis reported in this manuscript. G.L. and D.M. Classification of recurrence-free survival and tumor grade collaborated with us for developing accurate predictive models based on machine learning and feature engineering. Several machine learning algorithms, including support vector machine, random forest, and logistic regression with L1, L2, and ElasticNet regularization, were investigated to classify the dichotomized disease recurrence outcome based on a range of different cutoff times. For logistic ADDITIONAL INFORMATION 36 37 regression, algorithm solvers including Liblinear (L1 and L2), Saga (L1), Supplementary information accompanies the paper on the npj Breast Cancer 38 39 40 Lbfgs (L2), Newtoncg (L2), and Sag (L2) were explored. All radiomic website (https://doi.org/10.1038/s41523-018-0078-2). features were normalized to a standard z-score prior to any model training. The predictive performance of the classifier methods was quantified using Competing interests: The authors declare no competing interests. the area under receiver operator characteristic curve (AUC). The model parameters were optimized using stratified nested cross-validation (CV), Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims with 3-fold inner and outer cross validation repeated 10 times. The nested in published maps and institutional affiliations. cross-validation approach repeatedly splits the data into training, validation, and testing sets in order to avoid potential for over-fitting when estimating optimal tuning parameters and to provide unbiased REFERENCES estimation of the prediction performance. Stratification with respect to 1. Kumar, V. et al. Radiomics: the process and the challenges. Magn. Reson. Imaging label class was applied during the nested cross-validation such that the 30, 1234–1248 (2012). folds were made by preserving the proportion of samples for each label 2. Lambin, P. et al. Radiomics: extracting more information from medical images class. The mean and 95% confidence interval of the nested cross-validation using advanced feature analysis. Eur. J. Cancer 48, 441–446 (2012). AUCs (thresholding the logistic regression predicted probabilities) were 3. Aerts, H. J. W. L. et al. Decoding tumour phenotype by noninvasive imaging using reported over the 1000 repetitions using a bootstrap approach. All PET a quantitative radiomics approach. Nat. Commun. 5, 4006 (2014). and MR radiomic features were included in the nested cross-validation. In 4. Nicolasjilwan, M. et al. Addition of MR imaging features and genetic biomarkers predicting RSF, we reported ElasticNet logistic regression algorithm for the strengthens glioblastoma survival prediction in TCGA patients. J. Neuroradiol. 42, ease of interpretability. To examine the predictive power of the PET and 212–221 (2015). MR radiomic features, the features with the fitted coefficient >0 were 5. Segal, E. et al. Decoding global gene expression programs in liver cancer by tallied among 1000 repetitions of 3-fold outer cross-validation loop. The noninvasive imaging. Nat. Biotechnol. 25, 675–680 (2007). proportion of the times that a radiomic feature was selected out of 3000 6. Cook, G. J. R. et al. Are pretreatment 18F-FDG PET tumor textural features in CVs was ranked and the top 10 features were presented as the key features non–small cell lung cancer associated with response and survival after chemor- for predicting recurrence-free survival. In predicting tumor grade, we adiotherapy? J. Nucl. Med. 54,19–26 (2013). reported logistic regression with L2 regularization and Lbfgs solver. The 7. Coroller, T. P. et al. CT-based radiomic signature predicts distant metastasis in key predictors were determined by those with the |model fitted coefficient| lung adenocarcinoma. Radiother. Oncol. 114, 345–350 (2015). >0.01 and ranked according to the method described above. Cross- 8. Parmar, C. et al. Radiomic feature clusters and prognostic signatures specific for validation was implemented using Python (version 3.5.5), and machine lung and head and neck cancer. Sci. Rep. 5,1–10 (2015). learning algorithms used in this study were implemented in the Python 9. Lambin, P. et al. Radiomics: the bridge between medical imaging and persona- scikit-learn package. lized medicine. Nat. Rev. Clin. Oncol. 14, 749–762 (2017). 10. Win, T. et al. Tumor heterogeneity and permeability as measured on the CT Code availability component of PET/CT predict survival in patients with non–small cell lung cancer. All software custom-built for extracting radiomics from MR and PET Clin. Cancer Res. 3591–3600 (2013).https://doi.org/10.1158/1078-0432.CCR-12- images, data analysis, and cross validation is available on request from the 1307. corresponding author (Y.S.). 11. Xu, R., Kido, S. & Suga, K. Texture analysis on 18 F-FDG PET/CT images to dif- ferentiate malignant and benign bone and soft-tissue lesions. Ann. Nucl. Med. 926–935 (2014). https://doi.org/10.1007/s12149-014-0895-9. Data availability 12. Desseroit, M., Visvikis, D. & Tixier, F. Development of a nomogram combining The imaging data that support the findings of this are available on request. clinical staging with 18 F-FDG PET/CT image features in non-small-cell lung Please contact the following authors for specific image and clinical data cancer stage I – III. Eur. J. Nucl. Med. Mol. Imaging 1477–1485 (2016). https://doi. used in this study: Y. Seo for the whole-body PET/CT image and N.M. org/10.1007/s00259-016-3325-5. Hylton for the breast MR images. The imaging data are not publicly npj Breast Cancer (2018) 24 Published in partnership with the Breast Cancer Research Foundation Exploration of PET and MRI radiomic features for decoding breast Shih-ying Huang et al. 13. Vaidya, M. et al. Combined PET/CT image characteristics for radiotherapy tumor 31. Murtagh, F. & Legendre, P. Ward’ s hierarchical agglomerative clustering method: response in lung cancer. Radiother. Oncol. 102, 239–245 (2012). which algorithms Implement Ward’ s Criterion? J. Classif. 31, 274–295 (2014). 14. Vallières, M., Freeman, C. R., Skamene, S. R. & El Naqa, I. A radiomics model from 32. Hartigan, J. A. & Wong, M. A. A K-Means clustering algorithm. Appl. Stat. 28, 100 joint FDG-PET and MRI texture features for the prediction of lung metastases in (1979). soft-tissue sarcomas of the extremities. Phys. Med. Biol. 60, 5471–5496 (2015). 33. Kaufman, L., Rousseeuw, P. J. Finding groups in data: an introduction to cluster 15. Rakha, E. A. et al. Breast cancer prognostic classification in the molecular era: the analysis. (1990). role of histological grade. Breast Cancer Res. 12, 207 (2010). 34. Bergsma, W. A bias-correction for Cramer’s V and Tschuprow’sT. J. Korean Stat. 16. Li, H. et al. Quantitative MRI radiomics in the prediction of molecular classifica- Soc. 42, 323–328 (2013). tions of breast cancer subtypes in the TCGA/TCIA data set. NPJ Breast Cancer 2, 35. Wilkerson, M. D. & Hayes, D. N. ConsensusClusterPlus: a class discovery tool with 16012 (2016). confidence assessments and item tracking. Bioinformatics 26, 1572–1573 (2010). 17. Mukhtar, R. A. et al. Clinically meaningful tumor reduction rates vary by pre- 36. LIBLINEAR–A Library for Large Linear Classification. accessed online on July 25, chemotherapy mri phenotype and tumor subtype in the I-SPY 1 TRIAL (CALGB 2018. 150007/150012; ACRIN 6657). Ann. Surg. Oncol. 20, 3823–3830 (2013). 37. Defazio, A., Bach, F. & Lacoste-Julien, S. SAGA: A fast incremental gradient 18. Parmar, C., Grossmann, P., Bussink, J., Lambin, P. & Aerts, H. J. W. L. Machine method with support for non-strongly convex composite objectives. Adv. Neural learning methods for quantitative radiomic biomarkers. Sci. Rep. 5, 13087 (2015). Inform. Process. Syst. 1–15 (2014). arXiv:1407.0202. 19. Li, H. et al. MR imaging radiomics signatures for predicting the risk of breast 38. Liu, D. C. & Nocedal, J. On the limited memory BFGS method for large scale cancer recurrence as given by research versions of MammaPrint, Oncotype DX, optimization. Mathematical Programming. 45, 503–528 (1989). and PAM50 gene assays. Radiology 0, 152110 (2016). 39. Yu, H. F., Huang, F. L. & Lin, C. J. Dual coordinate descent methods for logistic 20. Bolouri, M. S. et al. Triple-negative and non–triple- negative invasive breast regression and maximum entropy models. Mach. Learn. 85,41–75 (2011). cancer: association between MR and fluorine 18 fluorodeoxyglucose PET Ima- 40. Schmidt, M. et al. Minimizing finite sums with the stochastic average gradient. ging. Radiology 269, 354–361 (2013). (2016), arXiv:1309.2388. 21. ACRIN. Protocol 6657. American College of Radiology Imaging Network https:// 41. Krstajic, D., Buturovic, L. J., Leahy, D. E. & Thomas, S. Cross-validation pitfalls when www.acrin.org/6657_protocol.aspx. selecting and assessing regression and classification models. J. Chemin-. 6,1–15 22. Partridge, S., Heumann, E., & Hylton, N. Semi-automated analysis for MRI of breast (2014). tumors. Stud. Health Technol. Inform. 62, 259–260 (1999). 42. Efron, B. Bootstrap methods: another look at the jackknife. Ann. Stat. 7,1–26 23. Partridge, S. C. et al. MRI measurements of breast tumor volume predict response (1979). to neoadjuvant chemotherapy and recurrence-free survival. Am. J. Roentgenol. 43. Buitinck, L. et al. API design for machine learning software: experiences from the 184(6), 1774–1781 (2005). scikit-learn project. 1–15 (2013). arXiv:1309.0238. 24. Klifa, C. et al. Quantification of breast tissue index from MR data using fuzzy clustering. Conf. Proc. Ieee. Eng. Med. Biol. Soc. 3, 1667–1670 (2004). 25. Fletcher, J. W. & Kinahan, P. E. PET/CT Standardized uptake values (SUVs) in Open Access This article is licensed under a Creative Commons clinical practice and assessing response to therapy. NIH Public Access 31, 496–505 Attribution 4.0 International License, which permits use, sharing, (2010). adaptation, distribution and reproduction in any medium or format, as long as you give 26. Chen, S. et al. Diagnostic classification of solitary pulmonary nodules using dual appropriate credit to the original author(s) and the source, provide a link to the Creative time 18F-FDG PET/CT image texture features in granuloma-endemic regions. Sci. Commons license, and indicate if changes were made. The images or other third party Rep. 7, 9370 (2017). material in this article are included in the article’s Creative Commons license, unless 27. Rahim, M. K. et al. Recent Trends in PET Image Interpretations Using Volumetric indicated otherwise in a credit line to the material. If material is not included in the and Texture-based Quantification Methods in NuclearOncology. Nucl. Med. Mol. article’s Creative Commons license and your intended use is not permitted by statutory Imaging 48,1–15 (2014). regulation or exceeds the permitted use, you will need to obtain permission directly 28. van Griethuysen, J. J. M. et al. Computational radiomics system to decode the from the copyright holder. To view a copy of this license, visit http://creativecommons. radiographic phenotype. Cancer Res. 77, e104–e107 (2017). org/licenses/by/4.0/. 29. American Joint Committee on Cancer. Breast cancer staging. 7th Ed. (2009) https://cancerstaging.org/references-tools/quickreferences/Documents/ © The Author(s) 2018 BreastMedium.pdf. 30. Monti, S., Tamayo, P., Mesirov, J. & Golub, T. Consensus clustering: a resampling- based method for class discovery and visualization of gene expression microarray data. Mach. Learn. 52,91–118 (2003). Published in partnership with the Breast Cancer Research Foundation npj Breast Cancer (2018) 24
npj Breast Cancer – Springer Journals
Published: Aug 16, 2018
You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
Access the full text.
Sign up today, get DeepDyve free for 14 days.
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.