Functional network analysis reveals an immune tolerance mechanism in cancer.

We present a technique to construct a simplification of a feature network which can be used for interactive data exploration, biological hypothesis generation, and the detection of communities or modules of cofunctional features. These are modules of features that are not necessarily correlated, but nevertheless exhibit common function in their network context as measured by similarity of relationships with neighboring features. In the case of genetic networks, traditional pathway analyses tend to assume that, ideally, all genes in a module exhibit very similar function, independent of relationships with other genes. The proposed technique explicitly relaxes this assumption by employing the comparison of relational profiles. For example, two genes which always activate a third gene are grouped together even if they never do so concurrently. They have common, but not identical, function. The comparison is driven by an average of a certain computationally efficient comparison metric between Gaussian mixture models. The method has its basis in the local connection structure of the network and the collection of joint distributions of the data associated with nodal neighborhoods. It is benchmarked on networks with known community structures. As the main application, we analyzed the gene regulatory network in lung adenocarcinoma, finding a cofunctional module of genes including the pregnancy-specific glycoproteins (PSGs). About 20% of patients with lung, breast, uterus, and colon cancer in The Cancer Genome Atlas (TCGA) have an elevated PSG+ signature, with associated poor group prognosis. In conjunction with previous results relating PSGs to tolerance in the immune system, these findings implicate the PSGs in a potential immune tolerance mechanism of cancers.

The maximum entropy principle for compositional data.

BACKGROUND: Compositional systems, represented as parts of some whole, are ubiquitous. They encompass the abundances of proteins in a cell, the distribution of organisms in nature, and the stoichiometry of the most basic chemical reactions. Thus, a central goal is to understand how such processes emerge from the behaviors of their components and their pairwise interactions. Such a study, however, is challenging for two key reasons. Firstly, such systems are complex and depend, often stochastically, on their constituent parts. Secondly, the data lie on a simplex which influences their correlations. RESULTS: To resolve both of these issues, we provide a general and data-driven modeling tool for compositional systems called Compositional Maximum Entropy (CME). By integrating the prior geometric structure of compositions with sample-specific information, CME infers the underlying multivariate relationships between the constituent components. We provide two proofs of principle. First, we measure the relative abundances of different bacteria and infer how they interact. Second, we show that our method outperforms a common alternative for the extraction of gene-gene interactions in triple-negative breast cancer. CONCLUSIONS: CME provides novel and biologically-intuitive insights and is promising as a comprehensive quantitative framework for compositional data.

PathCNN: interpretable convolutional neural networks for survival prediction and pathway analysis applied to glioblastoma.

MOTIVATION: Convolutional neural networks (CNNs) have achieved great success in the areas of image processing and computer vision, handling grid-structured inputs and efficiently capturing local dependencies through multiple levels of abstraction. However, a lack of interpretability remains a key barrier to the adoption of deep neural networks, particularly in predictive modeling of disease outcomes. Moreover, because biological array data are generally represented in a non-grid structured format, CNNs cannot be applied directly. RESULTS: To address these issues, we propose a novel method, called PathCNN, that constructs an interpretable CNN model on integrated multi-omics data using a newly defined pathway image. PathCNN showed promising predictive performance in differentiating between long-term survival (LTS) and non-LTS when applied to glioblastoma multiforme (GBM). The adoption of a visualization tool coupled with statistical analysis enabled the identification of plausible pathways associated with survival in GBM. In summary, PathCNN demonstrates that CNNs can be effectively applied to multi-omics data in an interpretable manner, resulting in promising predictive power while identifying key biological correlates of disease. AVAILABILITY AND IMPLEMENTATION: The source code is freely available at: https://github.com/mskspi/PathCNN.

Pan-Cancer Prediction of Cell-Line Drug Sensitivity Using Network-Based Methods.

The development of reliable predictive models for individual cancer cell lines to identify an optimal cancer drug is a crucial step to accelerate personalized medicine, but vast differences in cancer cell lines and drug characteristics make it quite challenging to develop predictive models that result in high predictive power and explain the similarity of cell lines or drugs. Our study proposes a novel network-based methodology that breaks the problem into smaller, more interpretable problems to improve the predictive power of anti-cancer drug responses in cell lines. For the drug-sensitivity study, we used the GDSC database for 915 cell lines and 200 drugs. The theory of optimal mass transport was first used to separately cluster cell lines and drugs, using gene-expression profiles and extensive cheminformatic drug features, represented in a form of data networks. To predict cell-line specific drug responses, random forest regression modeling was separately performed for each cell-line drug cluster pair. Post-modeling biological analysis was further performed to identify potential biological correlates associated with drug responses. The network-based clustering method resulted in 30 distinct cell-line drug cluster pairs. Predictive modeling on each cell-line-drug cluster outperformed alternative computational methods in predicting drug responses. We found that among the four drugs top-ranked with respect to prediction performance, three targeted the PI3K/mTOR signaling pathway. Predictive modeling on clustered subsets of cell lines and drugs improved the prediction accuracy of cell-line specific drug responses. Post-modeling analysis identified plausible biological processes associated with drug responses.

A machine learning model that classifies breast cancer pathologic complete response on MRI post-neoadjuvant chemotherapy.

BACKGROUND: For breast cancer patients undergoing neoadjuvant chemotherapy (NAC), pathologic complete response (pCR; no invasive or in situ) cannot be assessed non-invasively so all patients undergo surgery. The aim of our study was to develop and validate a radiomics classifier that classifies breast cancer pCR post-NAC on MRI prior to surgery. METHODS: This retrospective study included women treated with NAC for breast cancer from 2014 to 2016 with (1) pre- and post-NAC breast MRI and (2) post-NAC surgical pathology report assessing response. Automated radiomics analysis of pre- and post-NAC breast MRI involved image segmentation, radiomics feature extraction, feature pre-filtering, and classifier building through recursive feature elimination random forest (RFE-RF) machine learning. The RFE-RF classifier was trained with nested five-fold cross-validation using (a) radiomics only (model 1) and (b) radiomics and molecular subtype (model 2). Class imbalance was addressed using the synthetic minority oversampling technique. RESULTS: Two hundred seventy-three women with 278 invasive breast cancers were included; the training set consisted of 222 cancers (61 pCR, 161 no-pCR; mean age 51.8 years, SD 11.8), and the independent test set consisted of 56 cancers (13 pCR, 43 no-pCR; mean age 51.3 years, SD 11.8). There was no significant difference in pCR or molecular subtype between the training and test sets. Model 1 achieved a cross-validation AUROC of 0.72 (95% CI 0.64, 0.79) and a similarly accurate (P = 0.1) AUROC of 0.83 (95% CI 0.71, 0.94) in both the training and test sets. Model 2 achieved a cross-validation AUROC of 0.80 (95% CI 0.72, 0.87) and a similar (P = 0.9) AUROC of 0.78 (95% CI 0.62, 0.94) in both the training and test sets. CONCLUSIONS: This study validated a radiomics classifier combining radiomics with molecular subtypes that accurately classifies pCR on MRI post-NAC.

Optimal mass transport kinetic modeling for head and neck DCE-MRI: Initial analysis.

PURPOSE: Current state-of-the-art models for estimating the pharmacokinetic parameters do not account for intervoxel movement of the contrast agent (CA). We introduce an optimal mass transport (OMT) formulation that naturally handles intervoxel CA movement and distinguishes between advective and diffusive flows. METHOD: Ten patients with head and neck squamous cell carcinoma (HNSCC) were enrolled in the study between June 2014 and October 2015 and underwent DCE MRI imaging prior to beginning treatment. The CA tissue concentration information was taken as the input in the data-driven OMT model. The OMT approach was tested on HNSCC DCE data that provides quantitative information for forward flux ( ΦF ) and backward flux ( ΦB ). OMT-derived ΦF was compared with the volume transfer constant for CA, Ktrans , derived from the Extended Tofts Model (ETM). RESULTS: The OMT-derived flows showed a consistent jump in the CA diffusive behavior across the images in accordance with the known CA dynamics. The mean forward flux was 0.0082 ± 0.0091 ( min-1 ) whereas the mean advective component was 0.0052 ± 0.0086 ( min-1 ) in the HNSCC patients. The diffusive percentages in forward and backward flux ranged from 8.67% to 18.76% and 12.76% to 30.36%, respectively. The OMT model accounts for intervoxel CA movement and results show that the forward flux ( ΦF ) is comparable with the ETM-derived Ktrans . CONCLUSIONS: This is a novel data-driven study based on optimal mass transport principles applied to patient DCE imaging to analyze CA flow in HNSCC.

A theoretical investigation of adequate range uncertainty margins in proton treatment planning to preserve tumor control probability.

Background: Proton dose distributions are sensitive to range uncertainties, resulting in margins added to ensure adequate tumor control probability (TCP). We investigated the required margin and dose shape needed to ensure adequate TCP, for representative tumor cell distributions in the clinical target volume (CTV). Material and methods: A mechanistic tumor response model, validated for lung tumors, was used to estimate TCP. The tumor cell distribution ( ρ ) was assumed to decrease exponentially in the CTV with decay parameter λ toward the outer border ( xCTVmax ). It was investigated if a gradual dose fall-off could reduce the dose to normal tissues outside the CTV, while achieving adequate TCP. For various values of xCTVmax and λ, we derived adequate uniform dose margins ( m ), coupled to linear dose fall-off regions ( Δx, Δxnom=Δx-0.9 cm), that ensured TCP>TCPlimit, while delivering the least mean dose outside the CTV. To account for variabilities in patients and tumor types, variable probabilities ( p ) of finding tumor cells in the non-GTV part of the CTV for a given patient were also tested. Dose from a single beam or two opposing beams was simulated under the influence of a typical stopping power ratio uncertainty of 3.5%. Results: For large λ and xCTVmax, a dose distribution with a shallower dose fall-off ( Δx>0 ) was advantageous, and m could be smaller than xCTVmax. In the case of small xCTVmax values, however, a conventional dose distribution ( Δx=0 ) would generally perform better. For no CTV, m=0.4 cm in the case of two opposing beams, while it was 0.7 cm for a single beam, however, for two opposing beams Δx=1.2 cm ( Δxnom=0.3 cm), while it was zero for a single beam. Conclusion: The details of the underlying cancer cell distribution characteristics do impact the adequate dose arrangements, and for opposing beams a non-conventional dose distribution shape is often advantageous.

Dosimetric evaluation of an atlas-based synthetic CT generation approach for MR-only radiotherapy of pelvis anatomy.

PURPOSE: To investigate the potential of an atlas-based approach in generation of synthetic CT for pelvis anatomy. METHODS: Twenty-three matched pairs of computed tomography (CT) and magnetic resonance imaging (MRI) scans were selected from a pool of prostate cancer patients. All MR scans were preprocessed to reduce scanner- and patient-induced intensity inhomogeneities and to standardize their intensity histograms. Ten (training dataset) of 23 pairs were then utilized to construct the coregistered CT-MR atlas. The synthetic CT for a new patient is generated by appropriately weighting the deformed atlas of CT-MR onto the new patient MRI. The training dataset was used as an atlas to generate the synthetic CT for the rest of the patients (test dataset). The mean absolute error (MAE) between the deformed planning CT and synthetic CT was computed over the entire CT image, bone, fat, and muscle tissues. The original treatment plans were also recomputed on the new synthetic CTs and dose-volume histogram metrics were compared. The results were compared with a commercially available synthetic CT Software (MRCAT) that is routinely used in our clinic. RESULTS: MAE errors (±SD) between the deformed planning CT and our proposed synthetic CTs in the test dataset were 47 ± 5, 116 ± 12, 36 ± 6, and 47 ± 5 HU for the entire image, bone, fat, and muscle tissues respectively. The MAEs were 65 ± 5, 172 ± 9, 43 ± 7, and 42 ± 4 HU for the corresponding tissues in MRCAT CT. The dosimetric comparison showed consistent results for all plans using our synthetic CT, deformed planning CT and MRCAT CT. CONCLUSION: We investigated the potential of a multiatlas approach to generate synthetic CT images for the pelvis. Our results demonstrate excellent results in terms of HU value assignment compared to the original CT and dosimetric consistency.

Dual-input tracer kinetic modeling of dynamic contrast-enhanced MRI in thoracic malignancies.

Pulmonary perfusion with dynamic contrast-enhanced (DCE-) MRI is typically assessed using a single-input tracer kinetic model. Preliminary studies based on perfusion CT are indicating that dual-input perfusion modeling of lung tumors may be clinically valuable as lung tumors have a dual blood supply from the pulmonary and aortic system. This study aimed to investigate the feasibility of fitting dual-input tracer kinetic models to DCE-MRI datasets of thoracic malignancies, including malignant pleural mesothelioma (MPM) and nonsmall cell lung cancer (NSCLC), by comparing them to single-input (pulmonary or systemic arterial input) tracer kinetic models for the voxel-level analysis within the tumor with respect to goodness-of-fit statistics. Fifteen patients (five MPM, ten NSCLC) underwent DCE-MRI prior to radiotherapy. DCE-MRI data were analyzed using five different single- or dual-input tracer kinetic models: Tofts-Kety (TK), extended TK (ETK), two compartment exchange (2CX), adiabatic approximation to the tissue homogeneity (AATH) and distributed parameter (DP) models. The pulmonary blood flow (BF), blood volume (BV), mean transit time (MTT), permeability-surface area product (PS), fractional interstitial volume (vI ), and volume transfer constant (KTrans ) were calculated for both single- and dual-input models. The pulmonary arterial flow fraction (γ), pulmonary arterial blood flow (BFPA ) and systemic arterial blood flow (BFA ) were additionally calculated for only dual-input models. The competing models were ranked and their Akaike weights were calculated for each voxel according to corrected Akaike information criterion (cAIC). The optimal model was chosen based on the lowest cAIC value. In both types of tumors, all five dual-input models yielded lower cAIC values than their corresponding single-input models. The 2CX model was the best-fitted model and most optimal in describing tracer kinetic behavior to assess microvascular properties in both MPM and NSCLC. The dual-input 2CX-model-derived BFA was the most significant parameter in differentiating adenocarcinoma from squamous cell carcinoma histology for NSCLC patients.

Diffusion-weighted MRI of the lung at 3T evaluated using echo-planar-based and single-shot turbo spin-echo-based acquisition techniques for radiotherapy applications.

PURPOSE: To compare single-shot echo-planar (SS-EPI)-based and turbo spin-echo (SS-TSE)-based diffusion-weighted imaging (DWI) in Non-Small Cell Lung Cancer (NSCLC) patients and to characterize the distributions of apparent diffusion coefficient (ADC) values generated by the two techniques. METHODS: Ten NSCLC patients were enrolled in a prospective IRB-approved study to compare and optimize DWI using EPI and TSE-based techniques for radiotherapy planning. The imaging protocol included axial T2w, EPI-based DWI and TSE-based DWI on a 3 T Philips scanner. Both EPI-based and TSE-based DWI sequences used three b values (0, 400, and 800 s/mm2 ). The acquisition times for EPI-based and TSE-based DWI were 5 and 8 min, respectively. DW-MR images were manually coregistered with axial T2w images, and tumor volume contoured on T2w images were mapped onto the DWI scans. A pixel-by-pixel fit of tumor ADC was calculated based on monoexponential signal behavior. Tumor ADC mean, standard deviation, kurtosis, and skewness were calculated and compared between EPI and TSE-based DWI. Image distortion and ADC values between the two techniques were also quantified using fieldmap analysis and a NIST traceable ice-water diffusion phantom, respectively. RESULTS: The mean ADC for EPI and TSE-based DWI were 1.282 ± 0.42 × 10-3 and 1.211 ± 0.31 × 10-3  mm2 /s. The average skewness and kurtosis were 0.14 ± 0.4 and 2.43 ± 0.40 for DWI-EPI and -0.06 ± 0.69 and 2.89 ± 0.62 for DWI-TSE. Fieldmap analysis showed a mean distortion of 13.72 ± 8.12 mm for GTV for DWI-EPI and 0.61 ± 0.4 mm for DWI-TSE. ADC values obtained using the diffusion phantom for the two techniques were within 0.03 × 10-3  mm2 /s with respect to each other as well as the established values. CONCLUSIONS: Diffusion-weighted turbo spin-echo shows better geometrical accuracy compared to DWI-EPI. Mean ADC values were similar with both acquisitions but the shape of the histograms was different based on the skewness and kurtosis values. The impact of differences in respiratory technique on ADC values requires further investigation.

A literature mining-based approach for identification of cellular pathways associated with chemoresistance in cancer.

Chemoresistance is a major obstacle to the successful treatment of many human cancer types. Increasing evidence has revealed that chemoresistance involves many genes and multiple complex biological mechanisms including cancer stem cells, drug efflux mechanism, autophagy and epithelial-mesenchymal transition. Many studies have been conducted to investigate the possible molecular mechanisms of chemoresistance. However, understanding of the biological mechanisms in chemoresistance still remains limited. We surveyed the literature on chemoresistance-related genes and pathways of multiple cancer types. We then used a curated pathway database to investigate significant chemoresistance-related biological pathways. In addition, to investigate the importance of chemoresistance-related markers in protein-protein interaction networks identified using the curated database, we used a gene-ranking algorithm designed based on a graph-based scoring function in our previous study. Our comprehensive survey and analysis provide a systems biology-based overview of the underlying mechanisms of chemoresistance.

Automatic classification of prostate cancer Gleason scores from multiparametric magnetic resonance images.

Noninvasive, radiological image-based detection and stratification of Gleason patterns can impact clinical outcomes, treatment selection, and the determination of disease status at diagnosis without subjecting patients to surgical biopsies. We present machine learning-based automatic classification of prostate cancer aggressiveness by combining apparent diffusion coefficient (ADC) and T2-weighted (T2-w) MRI-based texture features. Our approach achieved reasonably accurate classification of Gleason scores (GS) 6(3 + 3) vs. ≥7 and 7(3 + 4) vs. 7(4 + 3) despite the presence of highly unbalanced samples by using two different sample augmentation techniques followed by feature selection-based classification. Our method distinguished between GS 6(3 + 3) and ≥7 cancers with 93% accuracy for cancers occurring in both peripheral (PZ) and transition (TZ) zones and 92% for cancers occurring in the PZ alone. Our approach distinguished the GS 7(3 + 4) from GS 7(4 + 3) with 92% accuracy for cancers occurring in both the PZ and TZ and with 93% for cancers occurring in the PZ alone. In comparison, a classifier using only the ADC mean achieved a top accuracy of 58% for distinguishing GS 6(3 + 3) vs. GS ≥7 for cancers occurring in PZ and TZ and 63% for cancers occurring in PZ alone. The same classifier achieved an accuracy of 59% for distinguishing GS 7(3 + 4) from GS 7(4 + 3) occurring in the PZ and TZ and 60% for cancers occurring in PZ alone. Separate analysis of the cancers occurring in TZ alone was not performed owing to the limited number of samples. Our results suggest that texture features derived from ADC and T2-w MRI together with sample augmentation can help to obtain reasonably accurate classification of Gleason patterns.

Intravoxel incoherent motion diffusion-weighted MRI during chemoradiation therapy to characterize and monitor treatment response in human papillomavirus head and neck squamous cell carcinoma.

PURPOSE: Characterize and monitor treatment response in human papillomavirus (HPV) head and neck squamous cell carcinoma (HNSCC) using intra-treatment (intra-TX) imaging metrics derived from intravoxel incoherent motion (IVIM) diffusion-weighted magnetic resonance imaging (DW-MRI). MATERIALS AND METHODS: Thirty-four (30 HPV positive [+] and 4 HPV negative [-]) HNSCC patients underwent a total of 136 MRI including multi-b value DW-MRI (pretreatment [pre-TX] and intra-TX weeks 1, 2, and 3) at 3.0 Tesla. All patients were treated with chemo-radiation therapy. Monoexponential (yielding apparent diffusion coefficient [ADC]) and bi-exponential (yielding perfusion fraction [f], diffusion [D], and pseudo-diffusion [D*] coefficients) fits were performed on a region of interest and voxel-by-voxel basis, on metastatic neck nodes. Response was assessed using RECISTv1.1. The relative percentage change in D, f, and D* between the pre- and intra-TX weeks were used for hierarchical clustering. A Wilcoxon rank-sum test was performed to assess the difference in metrics within and between the complete response (CR) and non-CR groups. RESULTS: The delta (Δ) change in volume (V)1wk-0wk for the CR group differed significantly (P = 0.016) from the non-CR group, while not for V2wk-0wk and V3wk-0wk (P > 0.05). The mean increase in ΔD3wk-0wk for the CR group was significantly higher (P = 0.017) than the non-CR group. ADC and D showed an increasing trend at each intra-TX week when compared with pre-TX in CR group (P < 0.003). Hierarchical clustering demonstrated the existence of clusters in HPV + patients. CONCLUSION: After appropriate validation in a larger population, these IVIM imaging metrics may be useful for individualized treatment in HNSCC patients. LEVEL OF EVIDENCE: 2 J. Magn. Reson. Imaging 2017;45:1013-1023.

Spatial rectal dose/volume metrics predict patient-reported gastro-intestinal symptoms after radiotherapy for prostate cancer.

BACKGROUND: Gastro-intestinal (GI) toxicity after radiotherapy (RT) for prostate cancer reduces patient's quality of life. In this study, we explored associations between spatial rectal dose/volume metrics and patient-reported GI symptoms after RT for localized prostate cancer, and compared these with those of dose-surface/volume histogram (DSH/DVH) metrics. MATERIAL AND METHODS: Dose distributions and six GI symptoms (defecation urgency/emptying difficulties/fecal leakage, ≥Grade 2, median follow-up: 3.6 y) were extracted for 200 patients treated with image-guided RT in 2005-2007. Three hundred and nine metrics assessed from 2D rectal dose maps or DSHs/DVHs were subject to 50-times iterated five-fold cross-validated univariate and multivariate logistic regression analysis (UVA, MVA). Performance of the most frequently selected MVA models was evaluated by the area under the receiving-operating characteristics curve (AUC). RESULTS: The AUC increased for dose-map compared to DSH/DVH-based models (mean SD: 0.64 ± 0.03 vs. 0.61 ± 0.01), and significant relations were found for six versus four symptoms. Defecation urgency and faecal leakage were explained by high doses at the central/upper and central areas, respectively; while emptying difficulties were explained by longitudinal extensions of intermediate doses. CONCLUSIONS: Predictability of patient-reported GI toxicity increased using spatial metrics compared to DSH/DVH metrics. Novel associations were particularly identified for emptying difficulties using both approaches in which intermediate doses were emphasized.

Associations between volume changes and spatial dose metrics for the urinary bladder during local versus pelvic irradiation for prostate cancer.

BACKGROUND: Inter-fractional variation in urinary bladder volumes during the course of radiotherapy (RT) for prostate cancer causes deviations between planned and delivered doses. This study compared planned versus daily cone-beam CT (CBCT)-based spatial bladder dose distributions, for prostate cancer patients receiving local prostate treatment (local treatment) versus prostate including pelvic lymph node irradiation (pelvic treatment). MATERIAL AND METHODS: Twenty-seven patients (N = 15 local treatment; N = 12 pelvic treatment) were treated using daily image-guided RT (1.8 Gy@43-45 fx), adhering to a full bladder/empty rectum protocol. For each patient, 9-10 CBCTs were registered to the planning CT, using the clinically applied translations. The urinary bladder was manually segmented on each CBCT, 3 mm inner shells were generated, and semi and quadrant sectors were created using axial/coronal cuts. Planned and delivered DVH metrics were compared across patients and between the two groups of treatment (t-test, p < .05; Holm-Bonferroni correction). Associations between bladder volume variations and the dose-volume histograms (DVH) of the bladder and its sectors were evaluated (Spearman's rank correlation coefficient, rs). RESULTS: Bladder volumes varied considerably during RT (coefficient of variation: 16-58%). The population-averaged planned and delivered DVH metrics were not significantly different at any dose level. Larger treatment bladder volumes resulted in increased absolute volume of the posterior/inferior bladder sector receiving intermediate-high doses, in both groups. The superior bladder sector received less dose with larger bladder volumes for local treatments (rs ± SD: -0.47 ± 0.32), but larger doses for pelvic treatments (rs ± SD: 0.74 ± 0.24). CONCLUSIONS: Substantial bladder volume changes during the treatment course occurred even though patients were treated under a full bladder/daily image-guided protocol. Larger bladder volumes resulted in less bladder wall spared at the posterior-inferior sector, regardless the treatment received. Contrary, larger bladder volumes meant larger delivered doses to the superior bladder sector for pelvic RT but smaller doses for local treatments.

Investigating the Robustness Neighborhood Gray Tone Difference Matrix and Gray Level Co-occurrence Matrix Radiomic Features on Clinical Computed Tomography Systems Using Anthropomorphic Phantoms: Evidence From a Multivendor Study.

OBJECTIVE: The aim of this study was to determine if optimized imaging protocols across multiple computed tomography (CT) vendors could result in reproducible radiomic features calculated from an anthropomorphic phantom. METHODS: Materials with varying degrees of heterogeneity were placed throughout the lungs of the phantom. Twenty scans of the phantom were acquired on 3 CT manufacturers with chest CT protocols that had optimized protocol parameters. Scans were reconstructed using vendor-specific standards and lung kernels. The concordance correlation coefficient (CCC) was used to calculate reproducibility between features. For features with high CCC values, Bland-Altman analysis was also used to quantify agreement. RESULTS: The mean Hounsfield unit (HU) was 32.93 HU (141.7 to -26.5 HU) for the rubber insert and 347.2 HU (-320.9 to -347.7 HU) for the wood insert. Low CCC values of less than 0.9 were calculated for all features across all scans. CONCLUSIONS: Radiomic features that are derived from the spatial distribution of voxel intensities should be particularly scrutinized for reproducibility in a multivendor environment.

IMRT QA using machine learning: A multi-institutional validation.

PURPOSE: To validate a machine learning approach to Virtual intensity-modulated radiation therapy (IMRT) quality assurance (QA) for accurately predicting gamma passing rates using different measurement approaches at different institutions. METHODS: A Virtual IMRT QA framework was previously developed using a machine learning algorithm based on 498 IMRT plans, in which QA measurements were performed using diode-array detectors and a 3%local/3 mm with 10% threshold at Institution 1. An independent set of 139 IMRT measurements from a different institution, Institution 2, with QA data based on portal dosimetry using the same gamma index, was used to test the mathematical framework. Only pixels with ≥10% of the maximum calibrated units (CU) or dose were included in the comparison. Plans were characterized by 90 different complexity metrics. A weighted poison regression with Lasso regularization was trained to predict passing rates using the complexity metrics as input. RESULTS: The methodology predicted passing rates within 3% accuracy for all composite plans measured using diode-array detectors at Institution 1, and within 3.5% for 120 of 139 plans using portal dosimetry measurements performed on a per-beam basis at Institution 2. The remaining measurements (19) had large areas of low CU, where portal dosimetry has a larger disagreement with the calculated dose and as such, the failure was expected. These beams need further modeling in the treatment planning system to correct the under-response in low-dose regions. Important features selected by Lasso to predict gamma passing rates were as follows: complete irradiated area outline (CIAO), jaw position, fraction of MLC leafs with gaps smaller than 20 or 5 mm, fraction of area receiving less than 50% of the total CU, fraction of the area receiving dose from penumbra, weighted average irregularity factor, and duty cycle. CONCLUSIONS: We have demonstrated that Virtual IMRT QA can predict passing rates using different measurement techniques and across multiple institutions. Prediction of QA passing rates can have profound implications on the current IMRT process.

Multi-institutional validation of a novel textural analysis tool for preoperative stratification of suspected thyroid tumors on diffusion-weighted MRI.

PURPOSE: Ultrasound-guided fine needle aspirate cytology fails to diagnose many malignant thyroid nodules; consequently, patients may undergo diagnostic lobectomy. This study assessed whether textural analysis (TA) could noninvasively stratify thyroid nodules accurately using diffusion-weighted MRI (DW-MRI). METHODS: This multi-institutional study examined 3T DW-MRI images obtained with spin echo echo planar imaging sequences. The training data set included 26 patients from Cambridge, United Kingdom, and the test data set included 18 thyroid cancer patients from Memorial Sloan Kettering Cancer Center (New York, New York, USA). Apparent diffusion coefficients (ADCs) were compared over regions of interest (ROIs) defined on thyroid nodules. TA, linear discriminant analysis (LDA), and feature reduction were performed using the 21 MaZda-generated texture parameters that best distinguished benign and malignant ROIs. RESULTS: Training data set mean ADC values were significantly different for benign and malignant nodules (P = 0.02) with a sensitivity and specificity of 70% and 63%, respectively, and a receiver operator characteristic (ROC) area under the curve (AUC) of 0.73. The LDA model of the top 21 textural features correctly classified 89/94 DW-MRI ROIs with 92% sensitivity, 96% specificity, and an AUC of 0.97. This algorithm correctly classified 16/18 (89%) patients in the independently obtained test set of thyroid DW-MRI scans. CONCLUSION: TA classifies thyroid nodules with high sensitivity and specificity on multi-institutional DW-MRI data sets. This method requires further validation in a larger prospective study. Magnetic Resonance in Medicine published by Wiley Periodicals, Inc. on behalf of International Society for Magnetic Resonance.

Breast cancer molecular subtype classifier that incorporates MRI features.

PURPOSE: To use features extracted from magnetic resonance (MR) images and a machine-learning method to assist in differentiating breast cancer molecular subtypes. MATERIALS AND METHODS: This retrospective Health Insurance Portability and Accountability Act (HIPAA)-compliant study received Institutional Review Board (IRB) approval. We identified 178 breast cancer patients between 2006-2011 with: 1) ERPR + (n = 95, 53.4%), ERPR-/HER2 + (n = 35, 19.6%), or triple negative (TN, n = 48, 27.0%) invasive ductal carcinoma (IDC), and 2) preoperative breast MRI at 1.5T or 3.0T. Shape, texture, and histogram-based features were extracted from each tumor contoured on pre- and three postcontrast MR images using in-house software. Clinical and pathologic features were also collected. Machine-learning-based (support vector machines) models were used to identify significant imaging features and to build models that predict IDC subtype. Leave-one-out cross-validation (LOOCV) was used to avoid model overfitting. Statistical significance was determined using the Kruskal-Wallis test. RESULTS: Each support vector machine fit in the LOOCV process generated a model with varying features. Eleven out of the top 20 ranked features were significantly different between IDC subtypes with P < 0.05. When the top nine pathologic and imaging features were incorporated, the predictive model distinguished IDC subtypes with an overall accuracy on LOOCV of 83.4%. The combined pathologic and imaging model's accuracy for each subtype was 89.2% (ERPR+), 63.6% (ERPR-/HER2+), and 82.5% (TN). When only the top nine imaging features were incorporated, the predictive model distinguished IDC subtypes with an overall accuracy on LOOCV of 71.2%. The combined pathologic and imaging model's accuracy for each subtype was 69.9% (ERPR+), 62.9% (ERPR-/HER2+), and 81.0% (TN). CONCLUSION: We developed a machine-learning-based predictive model using features extracted from MRI that can distinguish IDC subtypes with significant predictive power. J. Magn. Reson. Imaging 2016;44:122-129.