The Quantitative Imaging Network The National Cancer Institute Then: 1939 And Now: 2016

Our research program continues to focus on the development and clinical implementation of quantitative breast MRI for assessing response to treatment. The goal of our QIN research is to develop accurate and reliable imaging techniques that can be applied as diagnostic tools to individualize treatments for breast cancer patients. Over the past year, our QIN efforts focused on 1) evaluation of advanced DWI approaches (high-spatial resolution DWI and DTI); 2) beginning dedicated breast PET studies in I-SPY 2; 3) providing curated image data and outcomes from I-SPY 1 for public data-sharing and 4) leading a QIN Grand Challenge to identify high performing imaging biomarkers of response.

We continue to evaluate high-spatial resolution (HR) DWI developed in collaboration with industrial partner General Electric for evaluating breast cancer response to neoadjuvant chemotherapy (NAC). We found that early changes in tumor HR ADC metrics were associated with pathologic complete response (pCR), with a trend of increasing area under the receiver operating curve (AUC) with decreasing ADC percentiles. We further investigated high-resolution diffusion-tensor imaging (DTI) metrics (fractional anisotropy (FA) and ADC) as predictors and found that early percent changes in tumor eigenvalues and ADC were the strongest DTI-derived predictors of pCR. These results were published in the QIN Special Issue of the journal Tomography (December 2016).

We recently began clinical studies to evaluate dedicated breast (db) PET in combination with DCE-MRI for assessing early treatment response. We believe dbPET, versus whole-body PET, represents an economically and clinically-feasible approach to utilization of PET in the NAC setting. dbPET may be more sensitive for evaluating early treatment response revealing functional changes that precede anatomic changes on MRI and offering greater specificity for characterizing treatment-specific molecular pathways. Early results in I-SPY 2 suggest unique and complementary imaging patterns between MRI and dbPET and among tumor subtypes defined by hormone receptor (HR) and HER2 status.

Over the past year we worked with The Cancer Imaging Archive (TCIA) to provide DCE- and DWI- breast imaging collections for public access, including the fully curated I-SPY 1 data set which became publicly available on September 1, 2016. Additional collections established on TCIA included a 64-patient pilot study for FTV evaluation of neoadjuvant treatment response and a 13-subject collection for use in the QIN ADC Mapping Challenge led by Dr. Newitt. The ADC Mapping Challenge results will be presented at the 2017 Annual Meeting of the ISMRM.

A QIN-sponsored Challenge to evaluate Breast MRI Metrics of Response (BMMR) led by Dr. Hylton opened in May 2016 and continued through October 2016. The objective of the BMMR Challenge was prediction of recurrence free survival time (RFS) for patients with invasive breast cancer undergoing NAC, utilizing serial DCE-MRI studies taken over the course of therapy. This was the first QIN challenge to be performed under the new QIN guidelines for Challenges and

Collaborative Projects (CCPs). The challenge was managed in collaboration with Dr. Jayashree Kalpathy-Cramer through the QINLabs website. Three QIN groups (U. Chicago (M. Giger), Moffitt Hospital (J. Drukteinis) and MGH (J. Kalpathy-Cramer)), and one non-QIN group (U. Pennsylvania (D. Kontos)) submitted results for evaluation. The BMMR Challenge results are currently being prepared for publication and will be presented at the 2017 QIN face-to-face meeting.

Shutter-Speed Model DCE-MRI for Assessment of Response to Cancer Therapy

Alina Tudorica, Xin Li, Karen Oh, Brooke Beckett, Stephen Chui, Kathleen Kemmer, Megan Troxell, Arpana Naik, Fred Loney, Charles Springer, Jayashree Kalpathy-Cramer, Shannon McWeeney, Christopher Ryan, Wei Huang

Oregon Health & Science University, Portland, Oregon

Introduction: The major goal of the OHSU QIN project is to evaluate the Shutter-Speed Model (SSM) DCE-MRI method for prediction and assessment of cancer response to therapy. Within the context of therapy response assessment, the SSM is compared with the more commonly used Tofts model (TM) for pharmacokinetic (PK) analysis of DCE-MRI data, with the only difference between the two models being that the former accounts for the finite inter-compartmental water exchange kinetics.

Methods: To assess breast cancer response to neoadjuvant chemotherapy (NACT), DCE-MRI scans were performed pre-NACT, after one NACT cycle, at mid-point of NACT, and after NACT completion. To assess soft-tissue sarcoma response to preoperative chemoradiotherapy, DCE-MRI scans were performed pre-therapy, after one cycle of chemotherapy, and after completion of chemoradiotherapy.
Pathologic response endpoints for breast cancer include pCR (pathologic complete response) vs. non-pCR and RCB (residual cancer burden) determined from the post-NACT resection specimens. Based on necrosis percentage (NP) of surgical specimens, soft tissue sarcoma was classified as optimal (NP ≥ 95%) or sub-optimal (NP < 95%) responder.
Results: Breast Cancer: it was found from 47 patients that the % changes of several PK parameters after the first NACT cycle were good (ROC AUC > 0.8) early predictors of response, with parameters of both PK models performing equally well. However, after first cycle and at NACT mid-point, RECIST LD (longest diameter) and its percent change remained poor (C < 0.7) predictors of response. Compared to non-pCRs, the pCRs had considerably larger decrease in K^trans and increase in τ_i (mean intracellular water lifetime – SSM-only parameter) after the first NACT cycle. After NACT completion, K^trans and k_ep of both models and RECIST LD were positively correlated, while τ_i was negatively associated with RCB. 3D texture features were extracted from parametric maps of PK parameters and semi-quantitative DCE-MRI parameters. Texture features from quantitative PK parameters were found to be more useful than those from semi-quantitative metrics for early prediction of breast cancer response to NACT, while the features from the SSM PK parameters were superior to the TM counterparts for prediction of response.
Soft tissue sarcoma: it was found from 22 patients that after one chemotherapy cycle, K^trans and its percent change were good early predictors of optimal vs. sub-optimal pathologic response with univariate logistic regression (ULR) C statistics values of 0.90 and 0.80, respectively, while RECIST LD percent change was just a fair predictor (C = 0.72). Post-therapy K^trans, v_e, and k_ep, not RECIST LD, were excellent (C > 0.90) markers of therapy response. Several DCE-MRI parameters pre-, during, and post-therapy, including K^trans, v_e, and k_ep, showed significant (P < 0.05) correlations with NP of resected tumor specimens.
Conclusion: Functional changes as measured by DCE-MRI are superior to changes in tumor size for early predictors of breast cancer and soft tissue sarcoma responses to preoperative therapy. PK parameters estimated with either the TM and SSM performed comparably well in early prediction of breast cancer NACT response and evaluation of RCB, suggesting that the systematic differences between the TM and SSM PK parameter values may not affect predictive capabilities. However, SSM analysis allows quantification of τ_i, a potential imaging biomarker of metabolic activity. The utility of τ_i is clearly demonstrated in early prediction of breast cancer therapy response and assessment of RCB. Results from the sarcoma study suggest that a functional imaging study such as DCE-MRI following preoperative therapy may yield additional information that is potentially more useful than tumor size for surgical planning and subsequent management.

Quantification of hepatocellular carcinoma heterogeneity with multi-parametric magnetic resonance imaging
Stefanie Hectors PhD, Mathilde Wagner MD PhD, Octavia Bane PhD, Cecilia Besa MD, Sara Lewis MD, Romain Remark PhD, Nelson Chen BS, M. Isabel Fiel MD, Hongfa Zhu MD, Sacha Gnjatic PhD, Miriam Merad MD PhD,

Yujin Hoshida MD PhD, Bachir Taouli MD
Mount Sinai, New York, NY
Purpose: To quantify tumor heterogeneity in HCC using multi-parametric MRI (mpMRI), and to report preliminary data correlating quantitative MRI parameters with histopathology and gene expression in a subset of patients.
Methods: Thirty-two prospectively enrolled patients (M/F 26/6, mean age 59y) with HCC underwent mpMRI including DWI, blood-oxygenation-level-dependent (BOLD), tissue-oxygenation-level-dependent (TOLD) and DCE-MRI. The following parameter maps were generated: apparent diffusion coefficient ADC from DWI; arterial flow F_a, portal flow F_p, total flow F_t, arterial fraction ART, mean transit time MTT and distribution volume DV from DCE-MRI; R₂* (before and after O₂) and ΔR₂* from BOLD-MRI and R₁ (before and after O₂) and ΔR₁ from TOLD-MRI. Histopathology and gene expression analysis was performed in 14 patients. For histopathology, one paraffin-embedded section was used for sequential staining of CD31 (endothelial cells), CD68 (macrophages) and CD3 (T-cells). A separate slide was used for HIF1α (hypoxia) staining. A threshold-based segmentation method was implemented to determine stained tumor fractions for each of the markers. For gene expression analysis, the following HCC marker genes were profiled: liver specific Wnt target (GLUL), stemness markers (EPCAM, KRT19), early HCC markers (BIRC5, HSP70, LYVE1, EZH2), a pharmacological target currently under clinical testing (FGFR4), potentially targetable angiogenesis marker (VEGFA) and targetable immune checkpoints (CD274, PDCD1, CTLA4). Histogram characteristics [central tendency parameters (mean and median) and heterogeneity parameters (standard deviation, kurtosis and skewness)] of MRI parameters in HCC and liver parenchyma were compared using Wilcoxon signed-rank tests. Inter-tumor heterogeneity was assessed using the coefficient of variation between histogram features across tumors. Histogram data was correlated between MRI methods in all patients and with histopathology and gene expression profiles in 14 patients.

Results: Thirty-nine HCC lesions were assessed size 4.4±3.3 cm). HCCs exhibited significantly higher intra-tissue heterogeneity compared to liver with all MRI methods (p<0.042). Inter-tumor heterogeneity was significantly higher for kurtosis and skewness vs. mean parameters (p<0.001).

While there were significant correlations for central tendency parameters between different MRI methods and with histopathology and gene expression, additional complementary correlations between BOLD and DCE-MRI and with histopathologic hypoxia marker HIF1α and gene expression of GLUL, EPCAM, KRT19, FGFR4 and PDCD1 were seen for the heterogeneity parameters (Figure 1).
Conclusion: Histogram analysis combining central tendency and heterogeneity parameters is promising for noninvasive HCC characterization on the functional, immunohistochemical and genomics level.
Figure 1. Heatmap showing significant correlations between mpMRI features (A), between mpMRI features and histopathology (B) and between mpMRI features and gene expression (C), showing complementarity of central tendency and heterogeneity parameters.

Staging and monitoring of treatment response of bladder cancer

by using CDSS-S and CDSS-T decision support systems
Lubomir Hadjiyski, PhD¹, Heang-Ping Chan, PhD¹, Kenny Cha, MSc¹, Alon Weizer, MD²,

Ajjai Alva, MD³, Elaine Caoili, MD¹, Richard H. Cohan, MD¹, Jun Wei, PhD¹, Scott Tomlins, MD, PhD^4
1Department of Radiology, ²Department of Urology, ³Department of Internal Medicine, Hematology-Oncology, ⁴Departments of Path. and Urology, Univ. of Michigan, Ann Arbor
Introduction: The goal of this project is to develop effective decision support tools that merge image-based radiomic and non-image-based biomarkers to assist radiologists and oncologists in assessment of bladder cancer stage and change as a result of treatment. Correct staging of bladder cancer is vital for minimizing the risk of under-treatment or over-treatment with neoadjuvant chemotherapy. It is also important to identify tumors that do not respond at an early stage of neoadjuvant therapy, allowing the patient a chance of alternative treatment.
Methods: To achieve this goal we continued the process of: (1) Collecting the database of multi-modality MR and CT exams of bladder cancers for development, training and testing of the decision support tools.

(2) Developing a Quantitative Image analysis tool for Bladder Cancer (QIBC) for estimation of gross tumor volume and other radiomic features in multimodality images. The 3D segmentation module of QIBC is critical for the entire project and has been improved based on descriptors of the local tissue texture and gray level properties inside and outside the lesion. We have also continued investigating the use of a deep learning convolution neural network for the task of segmenting the bladder wall [1] and the bladder lesions [2].

(3) Developing a computerized decision support system (CDSS-S) to assist clinicians in cancer staging. An objective decision support system that merges the information in a predictive model based on statistical outcomes of previous cases and machine learning may assist clinicians in making more accurate and consistent staging assessments. During the current time period of the project we have developed a CDSS-S to stage bladder cancer based on different machine learning techniques. Morphological and texture features were extracted from the segmented tumor. A linear discriminant analysis (LDA), a neural network (NN), a support vector machine (SVM), and a random forest (RAF) classifier were used to combine the features into a single score. The data set was split into Set 1 and Set 2 for two-fold cross validation.

(4) Developing a computerized decision support system (CDSS-T) to assist clinicians in evaluation of the change in the tumor characteristics as a result of neoadjuvant treatment. We evaluated three unique predictive models, which employ different fundamental design principles: 1) a pattern recognition method (DL-CNN), 2) a more deterministic radiomics-feature-based approach (RF-SL), and 3) a bridging method between the two that extracts features from image patterns (RF-ROI). DL-CNN: Regions of interests (ROIs) were extracted from within the segmented lesions from corresponding pre- and post-treatment scans of a patient and were paired together in multiple combinations to generate pre-post-treatment paired ROIs. RF-SL: A radiomics-feature-based analysis was applied to the segmented lesions (RF-SL) to build a classifier for the prediction of complete responders to chemotherapy. RF-ROI: Radiomic features from the paired ROIs (RF-ROI) were used to build a classifier for the prediction of complete responders to chemotherapy. An observer performance study with two experienced radiologists was performed independently, in which the radiologist estimated the likelihood of stage T0 after viewing each pre-post-treatment CTU pair.

Results: By using the texture features in the CDSS-S, the LDA classifier achieved an average test A_z of 0.90 (average of Set 1 and Set 2 test A_z). The average test A_z of the NN classifier was 0.91. The SVM classifier achieved an average test A_z of 0.90. The average test A_z of the RAF classifier was 0.93, respectively. The test A_z values for prediction of T0 disease after treatment with CDSS-T were 0.73 ± 0.08, 0.77 ± 0.08, 0.67 ± 0.08 for the DL-CNN, RF-SL, and RF-ROI methods, respectively. The two radiologists had A_z values of 0.76 ± 0.08 and 0.77 ± 0.07, respectively, on the test set.

Conclusions: The study shows the promise of the CDSS–S for assessing bladder cancer stage. This study also indicates the potential of using DL-CNN and radiomic features obtained by QIBC, to assist in assessment of treatment response.
[1] Cha KH, Hadjiiski L, Samala RK, Chan H-P, Caoili EM, Cohan RH, “Urinary Bladder Segmentation in CT Urography using Deep-Learning Convolutional Neural Network and Level Sets.” Medical Physics 2016; 43: 1882-96. PMCID: PMC4808067.
[2] Cha KH, Hadjiiski L, Samala RK, Chan H-P, Cohan RH, Caoili EM, Paramagul C, Alva A, Weizer AZ, “Bladder Cancer Segmentation in CT for Treatment Response Assessment: Application of Deep-Learning Convolution Neural Network – A Pilot Study.” Tomography 2016; 2: 421-8. PMCID: PMC5241049.

Advancing Quantification of Diffusion MRI for Oncologic Imaging

Thomas L. Chenevert, Dariya Malyarenko, Lauren Keith,

Craig J. Galban and Brian D. Ross
University of Michigan, School of Medicine, Ann Arbor, MI 48109

The overarching goal of our QIN project is to provide for standardized implementation and clinical validation of quantitative diffusion-weighted imaging (qDWI) techniques for cancer patients across multiple MRI systems in order to improve their use in multi-site cancer imaging trials. Project activities are aligned with three specific objectives involving collaborations with Imbio, LLC (industrial partner), the National Institute of Standards and Technology (NIST), Eastern Cooperative Oncology Group and the American College of Radiology Imaging Network (ECOG-ACRIN), Imaging and Radiation Oncology Core (IROC) and the RSNA Quantitative Imaging Biomarker Alliance (QIBA). This poster reviews progress on distinctive components of this research:

1. 
   To advance the use of qDWI for therapeutic response assessment for improving clinical management of cancer patients, we are developing a standardized software platform allowing for diffusion analysis. DWI metrics will be used to evaluate quantification of tumor diffusivity as an imaging response biomarker through histogram and voxel-based parametric response map (PRM) quantities. To derive predictive histogram metrics least sensitive to image co-registration errors, validation of outcome prediction for PRM versus histogram metrics was performed for 83 brain tumor patients using common set of tumor ROIs. Validated DWI metrics will next be integrated into precision medicine workflow to improve diagnostic and therapeutic response assessment protocol for cancer patients.
   

2. 
   To improve technical reproducibility of measured qDWI metrics, our team provided QA support for multi-site clinical trials within ECOG-ACRIN (breast ACRIN-6698 and ACRIN-6701; prostate ACRIN-6702) and IROC (brain NRG BN-001). Quantitative QA metrics derived from the ice-water diffusion phantom scans that were used for system certification and longitudinal performance evaluation was included in technical performance assessment by QIBA ADC profile of breast, brain and prostate. In collaboration with NIST, we continued development of in situ thermometer for the ambient temperature DWI phantom that spans clinically-relevant ADC range. Our relevant efforts focused on reduction of temperature calibration errors through improved probe chemical design and on automation of temperature measurements on clinical scanners using low-bandwidth EPI.
   
3. 
   To empirically characterize and retrospectively correct for non-uniform diffusion weighting bias, as a major source of technical variability in multi-site trials, our team has designed and lead the two-phase QIN collaborative project (CCP). As a result of this CCP, empiric bias correctors were derived and validated for six representative MRI gradient systems utilized by QIN centers. To streamline bias correction for qDWI clinical trials, automated correction tools were developed that recast the static (empiric) 3D correctors for arbitrary scan geometry according to DICOM. In parallel to empiric retrospective correction, our QIN team continued to lead the six-party academic-industrial partnership with three major MRI vendors and two other QIN centers to implement prospective correction of nonlinearity bias in diffusion weighting directly on clinical scanners.
   

Predictive Radiomics, Pathomics, and Radiology-Pathology Fusion Tools:

Applications to predicting treatment response in

Brain, Rectal, and Oropharyngeal Cancers

Satish Viswanath, Pallavi Tiwari, Anant Madabhushi

Case Western Reserve University

INTRODUCTION: Over one million patients undergo chemotherapy and radiotherapy as first-line treatment for cancer in the United States. After initial confirmatory diagnosis via biopsy and histopathological examination, patients undergo follow up imaging (MRI, CT, or PET) after treatment, to distinguish responders from non- and partial-responders. Currently, clinicians must either: (a) over-treat all patients with aggressive therapy (to minimize chance of failure), or (b) subject all patients to radical aggressive surgery (to ensure tumor is completely excised). Both these options are associated with significant issues relating to quality-of-life, patient care, and healthcare costs. In other words, there is a desperate clinical need for either (a) identifying patients at diagnosis who suffer more aggressive tumors and thus may not respond to first-line therapy, or (b) predicting response to treatment early in patients who are non-responsive to a given therapy.

For example, in a major challenge in management of over 60,000 brain tumor patients is distinguishing radiation necrosis (a benign complication of radiotherapy) from tumor recurrence on MRI; surgical intervention being the only reliable means of diagnosis. Similarly, in rectal cancers, over 10,000 patients undergo unnecessary mesorectal surgery because post-treatment imaging fails to accurately assess pathologic complete response. Finally, in the case of oropharyngeal cancers, while most p16+ tumors are typically biologically indolent, there is a significant minority of these patients who have aggressive disease and suffer distant metastasis and death. However, there are no clinical or molecular markers to distinguish the two subtypes at diagnosis, which could guide treatment escalation to ensure optimal patient outcome.

METHODOLOGY: We present several computational imaging tools recently developed by our group to address problems in prediction and response evaluation applied to brain, rectal, and oropharyngeal tumors. Based on routinely acquired imaging and pathology data, we have developed a variety of new radiomics (computerized feature analysis of radiographic images for disease characterization), pathomics (quantitative histomorphometric analysis of digitized pathology images), and radiology-pathology fusion (spatially mapping pathologic information and annotations onto imaging) methods.

RESULTS:

1. 
   Brain Tumors: We have developed new radiomics features specifically optimized for treatment response evaluation, such as CoLlAGe, which attempts to captures entropy of pixel-level gradient orientations. CoLlAGe yielded an accuracy of 82% in differentiating patients with radiation necrosis and recurrent tumor on a cohort of N=60 post-treatment brain MRI studies [1]. By comparison, state-of-the-art radiomics features such as cooccurrence, Gabor, run-length based features, and histogram of gradient orientations (HoG) yielded a significantly lower cross-validated accuracy of 60-70%.
   
2. 
   Rectal Cancers: Based on the availability of resection specimens, we have successfully constructed spatially fused, deeply annotated post-treatment datasets comprising ex vivo pathology and corresponding presurgical MRI for rectal cancers with < 2% error. Through this mapping, we found that radiomics features associated with different treatment effects (fibrosis, ulceration), residual disease, and non-cancerous rectal tissue were markedly different. In a cohort of N=32 rectal cancer patients, we have demonstrated 80-90% accuracy in distinguishing between different pathologic stages of response, based off a clustering evaluation of gradient directionality based radiomics features derived from post-chemoradiation T2w MRI [2].
   
3. 
   Oropharyngeal Cancers: A quantitative histomorpmetry based classifier trained using features of nuclear architecture yielded an accuracy of 75% via randomized cross validation in distinguishing between p16+ oropharyngeal cancers that do and do not progress [3]. Tumors that had worse outcome tended to have more complex nuclear architecture compared to tumors that had favorable outcome.
   

CONCLUSION: Our current results suggest that the use of computational imaging and machine learning tools could allow for discovery of sub-visual biomarkers for predicting outcome and response of brain, rectal and oropharyngeal cancers. Multi-site, independent validation of these radiomic and pathomic features is needed.

REFERENCES: [1] Prasanna, P et al, Nature Scientific Reports 2016

[2] Antunes, J, et al. ISMRM 2016

[3] Lewis et al. American Journal of Surgical Pathology 2014