The Quantitative Imaging Network The National Cancer Institute Then: 1939 And Now: 2016
Yüklə 4,95 Mb.
|
MRI Subgroup MRI data acquisition CCPs on accuracy of T1 mapping and validation of DWI bias correction Octavia Bane1,2, Dariya Malyarenko3, Lori L Arlinghaus4, Madhava Aryal5, Michael Boss6, Yue Cao5, Fiona Fennessy7, Stefanie Hectors1,2, Karl G. Helmer8, Wei Huang,9, Nola Hylton10, Michael A. Jacobs11, Jayashree Kalpathy-Cramer8, Kathryn Keenan6, Robert Mulkern7, David Newitt10, Stephen E. Russek6, Karl F. Stupic6, Mathilde Wagner1,2, Lisa Wilmes10, Thomas Yankeelov12, Yi-Fei Yen8, Thomas Chenevert3, Bachir Taouli1,2 1Translational and Molecular Imaging Institute, Icahn School of Medicine at Mount Sinai, 2Radiology, Icahn School of Medicine at Mount Sinai, 3Radiology, University of Michigan, 4Vanderbilt University Medical Center, 5Radiation Oncology, University of Michigan, 6National Institute of Standards and Technology, 7Radiology, Brigham and Women’s Hospital, 8Radiology, Massachusetts General Hospital, 9Radiology, Oregon Health and Science University, 10Radiology, University of California San Francisco, 11Radiology, Johns Hopkins University, 12Radiology, University of Texas at Austin Accuracy, repeatability and inter-platform reproducibility measurements of T1 quantification methods used for DCE-MRI: results from a multicenter phantom study Introduction: Accurate, repeatable and reproducible quantification of baseline tissue T1 is highly important for patient-specific, repeatable perfusion quantification from DCE-MRI studies. The objectives of this challenge project were to: 1) measure the accuracy and test-retest repeatability and 2) determine the inter-platform reproducibility of T1 quantification protocols or methods at 1.5T and 3.0T. Methods: A dedicated T1 phantom was imaged at 8 centers on 10 platforms (1.5T/3T: 2/8, from three vendors, GE, Philips and Siemens) with common inversion recovery spin echo (IR-SE) and variable flip angle (VFA) protocols, as well as 14 site-specific protocols [VFA, variable TR (VTR), proton density (PD) and Look-Locker IR]. Factors influencing accuracy with respect to reference NMR T1 measurements and repeatability were assessed using general linear mixed models. Inter-platform reproducibility with the common protocols was assessed by coefficient of variation (CV). Results: For the common IR-SE protocol, accuracy (error range across platforms 3.7%-9.0%) was influenced by the T1 solution vial (P<10-6), and test-retest repeatability (error 0.24%-9.06%) by scanner (P<10-6), while for the common VFA protocol, accuracy (error 7.1%-30.5%) was influenced by field strength (P=0.006), and repeatability (error 1.32%-25.5%) by scanner (P<0.0001). Among site-specific protocols, VFA with 2-3 flip angles and Look-Locker IR protocols had accuracy and repeatability errors <15%. Inter-platform reproducibility was higher for the common IR-SE (mean CV 1.5T/3T: 5.4%/4.7%) than the common VFA protocol (mean CV 1.5T/3T: 8.9%/15.3%). Conclusions: Based on our results, we recommend the use of an optimized VFA protocol for accurate T1 quantification, preferably at 1.5T. Validation of gradient non-linearity bias correction workflow for quantitative diffusion-weighted imaging in multicenter trials Introduction: The objective of this project was to demonstrate feasibility of centralized retrospective system-specific correction of gradient nonlinearity (GNL) bias for quantitative diffusion weighted imaging (DWI) across diverse scanners independent of scanned object, and therefore, applicable in multi-site clinical trials. Methods: Six representative MR scanner models were selected (two for each vendor: GE, Philips and Siemens.) Using corrector maps generated from gradient system characterization by ice-water phantom in the previous project phase, GNL bias correction was performed for test ADC measurements from an independent DWI phantom (room-temperature agar). The pre-computed three-dimensional GNL correctors were retrospectively applied to test DWI scans by the central analysis site. The correction was blinded to reference DWI of the agar phantom acquired by sites at magnet isocenter where GNL bias is negligible. The performance was evaluated from changes in ADC ROI histogram statistics before and after correction with respect to the unbiased reference ADC values provided by sites. Results: Both absolute error and non-uniformity of ADC map induced by GNL (median: 12%, range: -35% to +10%) were substantially reduced by correction (seven-fold in median and three-fold in range). Correction of systematic GNL bias resulted in two-fold decrease of technical variability across scanners (down to site temperature range). Conclusions: This work has demonstrated that centralized retrospective correction of GNL bias in diffusion weighting, obtained from one-time empiric characterization of system GNL, is warranted by the stability of gradient channel characteristics, is desired for substantial reduction of ADC map bias, and is clearly feasible for multi-center clinical trial setting. When not corrected, this technical bias both shifts and artificially broadens the corresponding ADC ROI histograms, and increases cross-system variability of the quantitative DWI metrics. The reduction of systematic ADC map errors using the proposed technology will have a positive impact on clinical trials that utilize quantitative parametric ADC maps in diagnostic and treatment response metrics. MRI Subgroup DSC & DCE Challenges and Collaborative Project Updates Kathleen M. Schmainda1, Melissa Prah1, Wei Huang2, Xin Li2, Laura Bell3, C.Chad Quarles3 and the MRI-Subgroup of Image Analysis Performance Metrics Work Group 1. Medical College of Wisconsin, 2. Oregon Health Science Institute, 3. Barrow Neurological Institute DSC-MRI Collaborative Project: The overall goal of this and future DSC projects is to reach consensus regarding the collection and post-processing of DSC-MRI data for the evaluation of patients with brain tumors. Co-registered T1w, DSC (acquired with contrast agent pre-load) and ROIs, including tumor, normal appearing cerebral cortex (NACC), normal appearing white matter, whole brain, and AIFs were made available from the Medical College of Wisconsin for 49 pathologically confirmed low (LG) and high-grade (HG) brain gliomas. A total of 8 participating sites generated DSC-MRI parameter maps, including relative cerebral blood volume (rCBV) and flow (CBF), using various SW (software) platforms (with or without leakage correction (LC) algorithms). With some overlap in SW platforms used across sites, 19 rCBV and 12 CBF parameter maps were produced for each glioma case. Using Lin’s Concordance Correlation Coefficient (CCC) we determined that measurements were both correlated and equal, with good (0.80 CCC < 0.9) or excellent (CCC 0.90) agreement observed in 19% or 75% of SW platforms for normalized rCBV (nRCBV) and in 35% or 59% of SW platforms for nCBF, respectively. Agreement was worse for NACC, where good or excellent agreement was 19% or 35% for nRCBV, and 24% or 18% for nCBF, respectively. Those that did not incorporate LC tended to show worse agreement for nRCBV. All SW platforms were able to distinguish LG and HG (P<0.0001), yet the thresholds for this distinction, based on ROC analysis ranged from 1.24-1.75 (nRCBV) or 1.26-2.26 (nCBF). However, it was determined that all SW platforms could distinguish LG and HG glioma with a sensitivity and specificity near 0.8 if a threshold of 1.45 is used, while a nCBF threshold of 1.84 could at best provide a sensitivity and specificity of only 0.64. Overall, these results show that there is substantial consistency of DSC post-processing among SW platforms and sites especially for nRCBV, and best with LC. DSC-MRI Digital Reference Object (DRO) Challenge: The purpose of this newly commenced challenge is two-fold: 1) to determine optimal post-processing methods while identifying post-processing steps that introduce variability to rCBV analysis and 2) to determine the variability in imaging scan parameters across sites and their influence on rCBV accuracy and precision. There are currently nine sites participating in this challenge. Using survey responses, site-specific DROs are being created and a standardized DRO will also be distributed for post-processing using site-established methods. To evaluate accuracy between ground truth rCBV and site-specific rCBV (due to both imaging parameters and post-processing methods) CCC will be calculated. It is anticipated that this challenge will establish the range of acquisition and post-processing strategies that yield high fidelity rCBV measures. DCE-MRI AIF (Arterial Input Function) Collaborative Project (Phase 2 & 3): In continuing phases of this multicenter data analysis challenge, the goal was to assess the effects of variations in arterial input function (AIF) quantification on estimation of DCE-MRI parameters. Nine QIN centers quantified AIFs from 11 pre-therapy prostate (Phase 2: P2) and from 7 pre- and post-therapy soft tissue sarcoma (Phase 3: P3) DCE-MRI data sets using center-specific methods. These AIFs and their reference-tissue-adjusted (RTA) variants were used by the managing center to perform pharmacokinetic (PK) data analysis using a shutter-speed model (P2) or standard Toft’s model (P3). All other variables including tumor ROI definition and pre-contrast T1 were kept constant to evaluate parameter variations caused by AIF variations only. Parameter variability was assessed using wCV and ICC. Variability in correlations of imaging parameters with clinical response end points was also assessed (P3 only). For both P2 and P3, AIF-caused variations were larger in Ktrans than ve and both were reduced when RTA AIFs were used. These variations were largely systematic, resulting in nearly unchanged parametric map patterns. Kep and τi (shutter-speed model only) were less sensitive to AIF uncertainty than Ktrans, suggesting that they may be more robust imaging biomarkers. Results from P3 demonstrate that there were no decreases in variations of PK parameter percent changes between therapy time-points, which suggests random errors in AIF quantification may occur in a longitudinal study. However, the uniform sign of correlations between imaging biomarkers and clinical response end points was encouraging, as this demonstrates the robustness of DCE-MRI for prediction of therapy response despite uncertainties in AIF determination. MRI Working Group DWI-Related Challenges and Collaborative Projects: Apparent Diffusion Coefficient (ADC) Mapping and Parametric ADC Map DICOM implementation David Newitt1, Dariya Malyarenko2, Andrey Fedorov3, and the MRI-Subgroup of Image Analysis Performance Metrics Work Group 1University of California, San Francisco, 2University of Michigan, 3Brigham and Women’s Hospital, Boston The objective of this collaborative project (CCP), focused on ‘Apparent Diffusion Coefficient (ADC) Mapping’, was to evaluate the feasibility of centralized analysis and the reproducibility of quantitative DWI maps from software platforms employed by institutions in the NCI Quantitative Imaging Network (QIN) for phantom and in vivo DWI acquisitions. Thirteen representative DICOM data sets from human (breast and liver cancer) DWI scans were de-identified to HIPAA standards and hosted on TCIA. Three additional DWI DICOM sets for a polyvinylpyrolidone (PVP) phantom (with six known ADC values) were provided by three QIBA RSNA sites and shared via NCIP-HUB. Each protocol included scans from 3 major MRI scanner manufacturers: Siemens, Philips, and GE Medical Systems. Fifteen analysis platforms from 11 QIN sites and 3 vendors were included in the study, based on IDL, MATLAB, 3D Slicer, Osirix, AFNI, and QIBAPhan1.3 DWI analysis implementations. File formats of the submitted maps included DICOM, NIFTI, NRRD and MATLAB. Centralized cross-site analysis was performed by UCSF (lead institution) using a common set of pre-defined ROIs. Observed biases for in vivo ADC measures were in the range of 2-3% from the group median with individual excursions to >5%. Outliers were traced to errors in the DICOM meta-data rather than differences in the ADC algorithms. Lower than 1% biases were observed for phantom ADC with general error trends consistent with in vivo results. Larger biases were common for more complex DWI models. This study highlighted the practical challenges of multi-center ADC analyses and the metric variation arising from choices of analysis algorithms, and showed that preservation of critical DICOM meta-data in DWI is essential to avoid significant errors. The ‘Parametric ADC Map DICOM Implementation’ CCP was launched as a supplemental effort for parent ‘ADC Mapping’ project, to assess current capability and provide future guidelines for generating DICOM-compliant parametric diffusion maps by the QIN community. Participating sites used a single multi-vendor DWI DICOM data set from a PVP phantom to generate parametric ADC maps. The resulting ADC map DICOM header metadata were evaluated by central analysis site (UMICH) and compared to general QIICR recommendations and DWI DICOM macro. The CCP analysis has confirmed that (a) current vendor DWI DICOMs deviate from standard, and (b) scanner-console (vendor-specific) ADC analysis software is not capable of parsing cross-vendor DWI DICOM. The CCP results show that no standard parametric ADC DICOM has been implemented by the community, and the source-image reference and ADC units/scale tags are mostly missing from the QIN site implementations. Furthermore, ADC fit parameters and models, deemed important for multi-site analysis in parent CCP, are missing both from the parametric map standard and site implementations. Guided by the current CCP findings and identified limitations, the final stage of the CCP will focus on inclusion of ADC fit parameters and models in parametric map standard and site DICOM dictionaries, as well as on evaluation of recently implemented ITCR solution (DCMQI tools) for uniform ADC DICOM generation across the QIN sites. Demonstration Abstracts Quantitative image analysis tools for assessing response in cancer therapy trials R. R. Beichel, E. J. Ulrich, C. Bauer, B. J. Smith, J. J. Sunderland, M. M. Graham,T. Casavant, M. Sonka, and J. M. Buatti University of Iowa To facilitate quantitative analysis of FDG PET scans in clinical trials, the interdisciplinary QIN team at the University of Iowa (NIH grant U01 CA140206) has developed open source extensions (software) for 3D Slicer. This effort was also supported by the Quantitative Image Informatics for Cancer Research (QIICR) project (U24 CA180918). The demonstration will provide an overview of developed freely available software, demonstrate installation, and provide practical application examples utilizing head and neck FDG PET image data. Specifically, the following tools will be presented. PET DICOM Extension - The PET DICOM Extension provides tools to import DICOM PET images into Slicer and performs Standardized Uptake Value (SUV) normalization. The software calculates Standardized Uptake Value (SUV) conversion factors and creates a corresponding DICOM Real World Value Mapping file. PET IndiC - The PET-IndiC Extension allows for fast semi-automated segmentation of regions of interest and calculation of quantitative indices. PET Liver Uptake Measurement - This tool allows automated measurement of liver uptake in SUV normalized FDG-18 whole-body PET scans. A detailed description and evaluation of the method can be found in [1]. PET Tumor Segmentation - This tool provides an Editor-Effect for semi-automated segmentation of tumors and hot lymph nodes in PET scans. A detailed description and evaluation of the method can be found in [2]. In addition, we will give a preview of fully-automated tools for the quantitative analysis of ACR/ACRIN-ECOG, SNMMI/CTN, and NEMA NU-2 image quality phantoms, which are currently in development and will help simplifying the process of PET image quality assessment. [1] C. Bauer, S. Sun, W. Sun, J. Otis, A. Wallace, B. J. Smith, J. J. Sunderland, M. M. Graham, M. Sonka, J. M. Buatti, and R. R. Beichel. Automated Measurement of Uptake in Cerebellum, Liver, and Aortic Arch in Full-body FDG PET/CT Scans. Medical Physics, 39(6), June 2012. [2] R. R. Beichel, M. Van Tol, E. J. Ulrich, C. Bauer, T. Chang, K. A. Plichta, B. J. Smith, J. J. Sunderland, M. M. Graham, M. Sonka, and J. M. Buatti, “Semiautomated segmentation of head and neck cancers in 18F-FDG PET scans: A just-enough-interaction approach,” Med Phys, 43(6), Jun 2016 Spectroscopic MRI Clinical Interface Gurbani, S. S., Shim, H. Emory University With recent advancements in high-resolution, high-speed whole brain 3D MR spectroscopy (MRS) sequences and analysis software, and combined with image co-registration algorithms, we have developed a whole-brain 3D high-resolution spectroscopic MRI (sMRI) pipeline and validated sMRI metrics with quantitative histological measures of tumor infiltration. To overcome key hurdles to widespread use of sMRI in clinical trials -- which include management of large data sets, the need for extensive image processing by a skilled MR spectroscopist, and communicating clinically useful information to physicians efficiently -- we developed a web-based application that centrally stores and analyzes sMRI and other MR-based images to be used in multisite clinical trial settings. The platform consists of two components: a server that handles all image processing, data storage, and inter-institution synchronization; and a client, which allows imaging scientists and clinicians at each site to access images via a web browser on a computer or tablet, without additional installed software. The clients are kept in sync with the server, enabling physicians to collaborate on the same patients/images. For multi-site/consortium clinical trials including advanced imaging modalities, it is necessary to develop a workflow for synchronized inter-institution analysis and a platform for real-time collaboration. We successfully developed such platform to integrate sMRI information with anatomical MR images. Observer Performance Evaluation of Bladder Cancer Treatment Response Assessment on CT scans using computerized decision support tool (CDSS-T) Lubomir Hadjiyski, PhD Heang-Ping Chan, PhD Kenny Cha, PhD Jun Wei, PhD Elaine M. Caoili, MD Richard H. Cohan, MD Alon Weizer, MD* Ajjai Alva, MD** Department of Radiology, *Department of Urology, **Department of Internal Medicine, Hematology-Oncology, University of Michigan, Ann Arbor We are developing a computerized decision support system (CDSS-T) to assist clinicians in evaluation of the change in the bladder tumor characteristics as a result of neoadjuvant treatment. An important component of CDSS-T is the graphical user interface built upon our general multi-modality, multi-disease MiViewer platform developed in CAD Research Laboratory. The MiViewer incorporates a number of image display and annotation tools, which facilitate collecting radiologists’ descriptors, annotating images, manual outlining lesions or organs, and editing the outlines. The MiViewer allows relatively easy expansion of its functionalities by incorporating new modules designed for specific applications. The MiViewer interfaces with our quantitative image analysis tool, referred to as multimodality Quantitative Image analysis tool for Bladder Cancer (QIBC), for evaluation of gross tumor volume (GTV) and radiomic characteristics on CT and MRI scans. A major component of the QIBC is a 3D segmentation module that provides multilevel processing, including advanced segmentation tools such as level sets and deep learning neural networks, and automatically estimates the tumor volume from the segmented tumor region and the pre- to post-treatment percent tumor volume change. For the CDSS-T, we have implemented an observer performance study module in the MiViewer to conduct observer studies for evaluation of radiologists’ and oncologists’ performance in cancer treatment response assessment with and without computerized decision support. The participants in the demonstration can have hands-on experience in using the computer-aided decision support system for evaluating bladder cancer treatment response: view the pre- and post-treatment CT Urography image pairs, estimate the likelihood that a case fully responded to treatment, view the CDSS-T’s likelihood score for treatment response, have the opportunity to change their assessment, and compare their assessment to clinical reference standard. QUANTITATIVE MRI OF GLIOBLASTOMA RESPONSE The QTIM suite of image analysis tools A. Beers, Y Yen, A. Mamonov, E. Gerstner B. Rosen, J. Kalpathy-Cramer Massachusetts General Hospital The goal of our project is to develop quantitative imaging tools to “facilitate clinical decision making,” particularly in the setting of treatment of glioblastoma (GBM) with antiangiogenic therapies. We will demonstrate a range of open source tools that our lab (the Quantitative Tumor Imaging at Martinos) has developed for cancer image analysis. The imaging methods being studied are Dynamic Contrast Enhance MRI (DCE-MRI), Dynamic Susceptibility Contrast MRI (DSC), and Diffusion Weighted MRI. The biomarkers being considered as useful in the assessment of early response to therapy include standard parameters such as Ktrans, rCBV, and ADC. In this demo, we will demonstrate the QTIM suite including: Tumor segmentation: we will demonstrate our Slicer based module for efficient tumor segmentation as well as our completely automatic deep learning based tool. Dynamic contrast enhanced MRI (DCE): we will demonstrate our python package for DCE analysis on the QIBA digital reference object as well as clinical cases. Dynamic susceptibility contrast enhanced MRI (DSC): we will demonstrate our Slicer based module for DSC using a model that incorporates leakage correction. Radiomics: we will demonstrate our radiomics pipeline on a set of radiomics phantoms as well as clinical data from the QIN lung feature challenge. Challenge/Pipeline platform where we will walk through
Yüklə 4,95 Mb. Dostları ilə paylaş:
|