The Quantitative Imaging Network The National Cancer Institute Then: 1939 And Now: 2016
Baseline and delta radiomic features improve prediction of lung cancer incidence from size-stratified nodules in the National Lung Screening Trial (NLST)
Yüklə 4,95 Mb.
|
Baseline and delta radiomic features improve prediction of lung cancer incidence from size-stratified nodules in the National Lung Screening Trial (NLST) Dmitry Cherezov1, Samuel Hawkins1, Dmitry Goldgof1, Lawrence O. Hall1, Yoganand Balagurunathan2, Robert J. Gillies2, Matthew B. Schabath3 1Department of Computer Sciences and Engineering, University of South Florida, Tampa, Florida; 2Department of Cancer Imaging and Metabolism, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida; 3Department of Cancer Epidemiology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida Background: Despite the many advantages and improvements associated with lung cancer screening, LDCT imaging is associated with a high rate of detection of indeterminate pulmonary nodules (IPNs) of which only a fraction actually developed into cancer. Clinical guidelines provide for the evaluation and follow-up of nodules; however, currently there are no validated clinical decision tools to predict lung cancer risk and probability of cancer development. Ideally, an efficient and accurate non-invasive approach should be developed as a clinical decision tool for radiologists and pulmonologists to better manage nodules, especially IPNs, in the lung cancer screening setting. As such, using LDCT images from the NLST, we extracted radiomic features from nodules of baseline positive screens (T0) and delta features from T0 to first follow-up (T1) and performed analyses to identify imaging features that predict incidence lung cancer. Since current national recommendations increased a positive scan threshold to a 6 mm longest diameter, we performed size-specific analyses using three size classes (< 6 mm [small], 6 to 16 mm [intermediate], and ≥ 16 mm [large]) and two size classes (< 6 mm [small] and ≥ 6 mm [large]). Methods: We extracted 219 features from T0 nodules, and 219 delta features which are the change in feature value from T0 to T1, including size, shape, location, and texture information. Nodules were identified for 160 cases that were diagnosed with incidence lung cancer at the first (T1) or second (T2) follow-up screen and for 307 nodule positive controls that had three consecutive positive screens (T0 to T2) not diagnosed as lung cancer. The cases and controls were split into a training cohort and a testing cohort and classifier models (Decision tree-J48, Rule Based Classier-JRIP, Naive Bayes, Support Vector Machine, Random Forests) stratified by size class were used to identify the most predictive feature sets. Results: The final models using three size classes and two size classes generally revealed modest improvements when baseline and delta features were considered vs. only baseline models. For the three size classes, the AUROC for small-sized nodules was 0.83 (95% CI 0.76 – 0.90) for baseline only radiomic features and 0.84 (95% CI 0.77 – 0.90) for baseline and delta features. For intermediate-sized nodules, the AUROC was 0.76 (95% CI 0.71 – 0.90) for baseline only radiomic features and 0.84 (95% CI 0.80 – 0.88 for baseline and delta features. For large-sized nodules, the AUROC was higher for baseline only radiomic features (AUROC = 0.86; 95% CI 0.75 – 0.91) compared to baseline and delta features (AUROC = 0.83; 95% CI 0.75 – 0.91). When the intermediate and large nodules were combined into a single group (≥ 6 mm), the AUROC for baseline only feature was 0.80 (95% CI 0.76 – 0.84) compared to an AUROC of 0.86 (95% CU 0.83 – 0.89) for baseline and delta features. Discussion: While LDCT for high-risk individuals has demonstrated unequivocally that early detection save lives, the current screening strategy comes at the cost of high false-positive rates, missing a large percentage of lung cancer cases, and limited clinical decision tools to manage nodules and to better estimate the risk of an individual to develop lung cancer in the future. To address the latter, we conducted a nested case-control analysis of the NLST to identify size-specific radiomic models that predict lung cancer incidence. Our analyses revealed modest improvements in predicting lung cancer incidence by combining baseline and delta radiomic features. Additionally, the nodule size-specific analyses revealed unique radiomic feature models for each size class. As such, these imaging features could be used in the clinical setting to improve current size-based nodule screening guidelines. Quantitative CT Imaging for Response Assessment when using Dose Reduction Methods Michael McNitt-Gray, PhD; Matthew Brown, PhD; Jonathan Goldin, MD, PhD; Grace Kim, PhD; John Hoffman, Wasil Wahi-Anwar, Nastaran Emaminejad, Angela Sultan Department of Radiological Sciences, David Geffen School of Medicine at UCLA Purpose: CT imaging is widely used for assessing response to therapy in clinical practice and in many clinical trials. Recent technical advances in CT have allowed improvements in imaging greater coverage, better resolution and lower radiation dose. However, the effects of these techniques on quantitative imaging methods have not been fully evaluated. The purpose of this work is to perform extensive evaluations of quantitative imaging metrics under a variety of acquisition and reconstruction conditions. Methods: Several quantitative imaging tasks have been identified that may be used in clinical trials or clinical practice including calculating properties of lung tumors such as volume, density and texture. Raw projection data from clinical imaging datasets have been acquired from our CT scanners and libraries to read and interpret these data have been developed. Using this raw data as the input, a high throughput pipeline has been created that allows creation of image datasets which represent a wide range of acquisition and reconstruction conditions; this includes a module that adds calibrated noise to simulate reduced dose acquisitions and a standalone reconstruction module that allows high throughput creation of image datasets at different reconstructed slice thicknesses or filter conditions. Using these datasets, we have performed initial evaluations on the effects of acquisition and reconstruction conditions on the specified quantitative imaging tasks (tumor volumes, densities and textures). Results: Results from a study of 33 “measurable” nodules were used to estimate inter-dose and inter-reconstruction-method reproducibility of tumor volumes. We compared the resulting distributions of proportional differences across dose and reconstruction methods by analyzing their means, standard deviations (SDs), and t-test and F-test results. The inter-dose reproducibility experiments gave mean differences ranging from −5.6% to −1.7% and SDs ranging from 6.3% to 9.9%. Analysis of representative nodules confirmed that reader variability appeared unaffected by dose or reconstruction method. This work demonstrated that lung-nodule volumetry was extremely robust to the radiation-dose level, down to the minimum scanner-supported dose settings. In addition, volumetry was robust to the reconstruction methods used in this study, which included both conventional filtered back projection and iterative methods Using the same nodules and water phantoms, we investigated the effects on density and texture based features. The water phantom results demonstrated substantial variability among feature values calculated across conditions, with the exception of histogram mean. Features calculated from lung nodules demonstrated similar results with histogram mean as the most robust feature. Surprisingly, histogram standard deviation and variance features were also quite robust. Some Spatial Gray Level Dependence Matrices (SGLDM or co-occurrence matrices) features were also quite robust across conditions, namely, diff. variance, sum variance, sum average, variance, and mean. Our conclusions were that the histogram mean – as expected - was the most robust feature in this study. The effects of acquisition and reconstruction conditions on GLCM features vary widely, though we did observe a trend that features involving summation of product between intensities and probabilities were more robust, barring a few exceptions. Overall, we observed that care should be taken to account for variation in density and texture features if a variety of dose and reconstruction conditions are used for the quantification of lung nodules in CT, otherwise changes in quantification results may be more reflective of changes due to acquisition and reconstruction conditions than in the nodule itself. Discussion: Through the experiments performed to date, tumor volume and density measures have been demonstrated to be robust to acquisition and reconstruction parameters, at least within reasonable limits. While some texture features have been shown to be somewhat stable across conditions, many of them vary substantially across these conditions. Future work will investigate strategies to mitigate these variations within and across scanner conditions including strict limits on allowed acquisition and reconstruction conditions, image correction or normalization (e.g. de-noising) methods as well as developing augmented datasets for machine learning approaches to overcome these limitations. Quantitative Volume and Density Response Assessment: Sarcoma and HCC as a Model Lawrence H Schwartz, M.D. and Binsheng Zhao, D.Sc. Columbia University Medical Center, New York City The goal of our research is to develop new response assessments for cancer treatment based on CT imaging of changes in tumor volume and necrosis fraction. Conventional RECIST criteria and cut-off values for response assessment are not evidence-based and may fail to detect tumor changes associated with clinical response to targeted, non-cytotoxic treatments. This study will seek a proof of concept using two types of tumors in which RECIST is known to correlate poorly with tumor response to treatment and clinical outcome. HCC is one of the most common malignancies worldwide, and sarcomas, though rare, carry the same molecular alterations as many other heterogeneous cancers and are the classic cancer studied in drug discovery. Since beginning our project in September 2011, we have been moving steadily through our specific aims. We have completed 1) collection and measurement of the two large clinical trial data, 2) correlative analysis of one trial data (sarcoma study) and a manuscript based on the study finding1, 3) optimization of our segmentation algorithms for solid tumors2,3, 4) investigation of variability in measuring tumor volume and volume change4, and 5) development of a prototype imaging system for tumor response assessment using an open source platform (Fig)5. We have been actively participated in the QIN collaborative projects, Challenges for PET and CT Lesion Segmentation, Lesion Change Analysis and Computation of Quantitative Image Features. We are now enhancing our necrosis segmentation method and working on analysis of the second trial data (HCC study). By the completion of our project, we plan to deliver the following: i) robust computer-aided segmentation and quantification algorithms for solid tumors, ii) insight into the magnitude of variability in measuring tumor volume (diameters as well), density and their changes, iii) a subset of CT images containing radiologists’ mark-up of tumor contours, and iv) new response metrics and criteria, based on tumor volume and necrosis fraction, for better assessing response to novel therapies. 
Publications
Lesion volume Estimation from PET without Requiring Segmentation Saeid Asgari1, Nóirín Duggan1, Hillgan Ma2, Anna Celler2, Francois Benard3, and Ghassan Hamarneh1 1 Medical Image Analysis Lab, Simon Fraser University, Burnaby, Canada 2 Medical Imaging Research Group, Department of Radiology, University of British Columbia, Vancouver, Canada 3 British Columbia Cancer Agency Research Centre, Vancouver, Canada Introduction: Tracking cancer progress and treatment efficacy requires quantifying tumour burden, which in turn benefits from estimating metrics based on total lesion volume, i.e. via the PET Response Criteria in Solid Tumors (PERCIST). This typically requires lesion delineation from a 3D volume, which can either be performed manually or using (highly) automated segmentation methods. The former approach is time-consuming and suffers from inter and intra-rater variability, whereas the latter can require computationally intensive processing, parameter tweaking and may result in ‘leaking’ or under-segmentation. Methods: We propose the first quantification method for lesion volume estimation that does not require segmentation. Our approach, at a high level, relies on training a machine learning system using a set of (image patch, volume value pairs). The image patch is a quick-to-draw, box-like 3D region of interest (ROI), i.e. not a delineation. In practical usage scenarios, given an ROI containing a lesion from a novel PET image, the trained model infers the volume (in mm^3) of the lesion within the ROI, without segmentation. We use regression Random Forests as the machine learning “engine”. Data: We trained our machine learning on a set of 140 phantom images, for which exact geometry and volume of lesion-like regions is available. The data included backgrounds of air, water and hot-background and had Signal to Background (SBR) values of 4:1 and 8:1, respectively. We tested our method on 52 clinical PET images from the Quantitative Imaging Network (QIN) Head and Neck collection in The Cancer Imaging Archive (TCIA). These images come from 3 scanners with different acquisition parameters and are accompanied by 3 expert segmentations for each lesion. Results: We used our method to estimate the volumes of 195 lesions from the 52 clinical PET images. Then we calculated the pairwise disagreement in volume estimations between the three experts and our method (i.e. disagreement between expert1 and expert2; expert2 and expert 3; expert1 and proposed method; etc.). Our preliminary results show that our method (i) achieves lower (by around 10%) disagreement with two of the experts (expert1 and expert3) than what the third expert (expert3) achieves, and (ii) has a level of disagreement with expert1 and expert2 similar (within 0.5%) to that of expert3. Conclusion: Quantifying lesion volume in clinical practice is important for managing cancer. In the past decades, numerous image segmentation-based methods were proposed in an attempt to automate this process. However, it remains difficult to choose a segmentation method and set its parameters. We presented the first machine learning-based model to directly estimate the tumor volume in PET images without applying segmentation. We show promising results in terms of close agreement with expert segmentation-based estimates. Correcting PET images of hypoxia for tissue transport properties Edward Taylor1,2, Jennifer Gottwald1,3, Ivan Yeung1,4, Harald Keller1,4, Neesha C. Dhani1,5, Michael Milosevic1,4, David W. Hedley1,3,5, David Jaffray1,2,3
Introduction: Compared to FDG, the binding rate of hypoxia-sensitive tracers such as fluoro-azomycinarabinoside (FAZA) is small, meaning that tracer uptake is sensitive to the presence of hypoxia as well as the transport properties (perfusion, diffusive transit time) of the imaged tissue. In order to accurately quantify hypoxia in solid tumours, PET images should be corrected for these. Methods: Dynamic PET time-activity curves were analyzed using a novel compartmental model in twenty patients with pancreatic ductal adenocarcinoma. A key transport quantity—the nonequilibrium partition coefficient—describing the voxel-scale ratio of unbound tracer concentrations in tissue and blood at short times after injection was identified in this model and used to interpret dynamic and static PET uptake metrics. Hypoxia was quantified by the fraction of voxels in which the FAZA binding rate exceeded a threshold value. Fig. 1 Left: Tumour-to-blood uptake ratio of FAZA in the voxels of a pancreatic tumour two hours after injection versus the trapping rate inferred from a compartmental model. Right: same quantity, but corrected for partitioning.
Fig. 2 Hypoxic fractions in twenty patients calculated from the binding rate and static uptake, ordered by partitioning. (=1 corresponds to zero partitioning). Results: Partitioning diminished the ability of static PET imaging to accurately quantify hypoxia: the tumour-to-blood uptake ratio (T/B) of FAZA after two hours was poorly correlated with the tracer trapping rate, a sum of the tracer binding rate due to hypoxia and the rate of diffusive transport across partitioned regions. Applying a theory-based partition correction, strong correlations were found; see Fig. 1. Hypoxic fractions calculated from the binding rate agreed with those calculated from static imaging (T/B > 1.2) when partitioning was small; otherwise, differences were significant (Fig. 2). Partitioning was well-correlated with hypo-attenuating features on non-contrast CT images, suggesting a physiological origin. Conclusions: Our results identified partitioning as a major challenge to accurate static PET-hypoxia quantification of pancreatic tumours. We used dynamic PET imaging to quantify hypoxia based on the tracer binding rate instead of uptake from a static scan to correct for its effects. The correlation between partitioning and static CT suggests that a combined static PET/CT biomarker may suffice to accomplish this in future. Impact of Experimental Settings on the Performance of PET radiomics in Predicting EGFR Mutation Status Stephen SF Yip1*, John Kim2, Chintan Parmar1, Elizabeth Huynh1, Raymond H. Mak1 and Hugo J.W.L. Aerts1,3 1Department of Radiation Oncology and 3Radiology, Dana-Farber Cancer Institute, Brigham and Women’s Hospital, and Harvard Medical School, Boston MA, USA 2Department of Radiology, Univ. of Michigan Health System, Ann Arbor MI, USA Purpose: A vital part of radiomic quantifications, is that medical images undergo image-processing methods, including resampling and intensity discretization, prior to the extraction of radiomic features. Various parameters can be chosen for these processing steps and the impact of these parameters on the performance of feature predictability is unclear. This study investigated how various settings affect the performance of radiomic features extracted from PET imaging in predicting somatic mutations. Methods: 348 non-small cell lung cancer (NSCLC) patients that underwent diagnostic PET images were tested for EGFR mutations. All PET images were resampled to uniform voxel sizes between 1x1x1 and 4x4x4 mm3 with four interpolation filters (nearest-neighbor, linear, cubic, and spline). The intensity histograms were discretized into equal widths between 0.1 and 0.5. In total, seven shape features were computed with 16 distinct experimental settings. Fifteen histogram features and 44 textural features were computed with 80 settings. The performance of the radiomic features in predicting EGFR mutations was assessed using the area under the receiver-operating-characteristic curve (AUC). The influence of different settings on feature predictability was quantified as the relative difference between the maximum and minimum AUC (δ). Results: The overall influence δOverall was 4.6% for shape, 12% for statistic, and 27% for textural features. Four shape, six statistic, twenty textural features were robust of experimental settings (δOverall<5%). Shape features were robust to filters and voxel size (δFilter≈δVoxel≈4%). Statistic features were robust to voxel size and bin width (δFilter, δVoxel<5%), but moderately sensitive to filters (δFilter=8.9%). Filters, voxel size, and bin width had a moderate impact on textural features (δ=5%–11%). Conclusion: While the predictability of 30 radiomic features were insensitive to the choice of the experimental settings, they need to be carefully chosen for all other features. Although all processing steps only had a small to moderate impact on textural features, their combined effect was substantial. Optimized settings that will maximize the predictive performance of individual features should be investigated in the future. Advanced PET/CT Imaging for Improving Clinical Trials Paul Kinahan, Hannah Linden, David Mankoff University of Washington and the Seattle Cancer Care Alliance, Seattle, WA University of Pennsylvania, Philadelphia, PA The goal of this project is to improve cancer clinical trials by enhancing the effectiveness of quantitative PET/CT imaging of tumor response. Phantom and software tools We have completed an evaluation of our PET/CT cross-calibration kit, which was designed in collaboration with RadQual. This is now available as a commercial product called the PET F-18 X-Cal System. The X-cal phantom and open-source measurement software package 'Xcaliper' are designed to allow the monitoring of biases in SUV values by measuring scanner and dose calibrator biases. Each source’s activity is known to ±2.5% with a 95% confidence level. In testing at multiple sites, per-site average recovery coefficients ranged from 0.907 to 0.983, with per-site standard deviations between 0.019 and 0.034. The 24 measurements overall had a mean of 0.944 ± 0.038. Dose calibrator recovery coefficients were 0.964 ± 0.033. Comparison of the pre-test PET scanner and dose calibrator biases did not show any correlations in the biases. These results were published in the QIN special issue of Tomography [1] showing longitudinal variations in bias at single-center and multi-center studies. The X-cal phantom kit was deployed in a QIN multi-center study that has completed analysis and is being submitted for publication.
Virtual Clinical Trials We developed a method called 'virtual clinical trials to evaluate variations in the PET imaging process to characterize the ability of static and dynamic metrics to measure breast cancer response to therapy in a clinical trial setting. We have competed and published three studies: Estimating the effect of FDG uptake time on lesion detectability in PET imaging of early stage breast cancer showing that delayed imaging improves detection [2], estimating the effects of uptake time variability on required sample size showing that variability in uptake time can double the needed number of patient studies in clinical trials [3], and comparing static versus dynamic PET imaging in measuring response to breast cancer therapy showing that as expected, dynamic imaging improves the correct discrimination of response [4]. [1] Byrd DW et al. Tomography 2(4):353-360, 2016. PMID: 28066807. PMCID: PMC5214172. [2] Wangerin KA, et al. Tomography 1:53-60, 2015. PMID: 26807443. PMCID: PMC4721230. [3] Kurland BF et al. J Nucl Med 57:226-230, 2016. PMID: 26493206. PMCID: PMC4749350. [4] Wangerin KA, et al. Phys Med Biol. (accepted) 2017. Development of advanced whole brain 3D spectroscopic MRI for the management of GBM patients Gurbani, S. S., Cordova, J.S., Hadjipanayis, C.G., Liang, Z., Cooper, L.A.D., Schreibmann, E., Shu, H.G., Olson, J.J., Holder, C.A., Shim, H. Emory University/Johns Hopkins Introduction: By combining the expertise and resources of Emory and Johns Hopkins Universities, our project is aim to develop clinically useful state-of-the-art spectroscopic MRI technology within the framework of a novel clinical trial. This method will allow quantitative and reproducible analyses of brain tumor therapeutic response, offering a precise and objective tool for longitudinal monitoring and inter-subject comparison of patients in clinical trials. Methods: Currently, contrast-enhanced T1-weighted MRI (CE-T1w) is used to monitor therapy response and define primary treatment volumes for surgery and radiation therapy (RT). However, contrast enhancement (blood-brain barrier breakdown) does not identify the tumor entirely, nor report therapy response reliably. MR spectroscopy has emerged as a molecular imaging modality which may better define the extent of active, viable tumor based on endogenous metabolites, without relying on perfusion/leakage/diffusion of an exogenously injected contrast medium. With recent advancements in high-resolution, high-speed whole brain 3D MRS sequences and analysis software, and combined with image co-registration algorithms, it is now possible to use MRS not only for monitoring therapy response but also for therapy planning. Careful validation of the state-of-the-art MRS metabolic maps with tissue histopathology was previously performed to ensure that these maps could be used clinically, because MRS abnormal lesions are often substantially larger than the area of enhancement on CE-T1w MRI. In recent years, we have developed a whole-brain 3D high-resolution spectroscopic MRI (sMRI) pipeline and validated sMRI metrics with quantitative histological measures of tumor infiltration. In August 2016, these data were published in Neuro-Oncology and featured on the cover. 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 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. Results: Our study demonstrated that sMRI, a non-invasive imaging technique that maps tumor metabolism without exogenous contrast media, could provide reliable biomarkers for the detection of the complete tumor profile and for monitoring therapy response. Heretofore, sMRI processing has been subjective and required hours of both computational and user time, hurdles that must be overcome for efficient and effective integration into daily clinical workflow. To address the aforementioned challenges, we have been developing software to push the technology towards clinical utilization, including automated, quantitative, expedient, and objective analysis methods of volumetric metabolic data. Current features of the web-based application include: importing images via Analyze, DICOM, and custom formats for sMRI; display of sMRI images overlaid on co-registered anatomical MRI; longitudinal comparison of sMRI scans; reporting quantitative tumor volumes; visualization of custom fields, such as the raw and fitted chemical shift spectra of individual voxels; algorithms for clustering voxels using mixture models; and exporting of tumor structures and associated images into clinical PACS or LINAC treatment systems. Conclusion: For multi-site/consortium clinical trials including advanced imaging modalities, it is necessary to develop a workflow for synchronized inter-institution analysis. We successfully developed a platform to integrate sMRI information with anatomical MR images, and will continue to adapt this platform for additional modalities and image processing algorithms. DSC-MRI in Brain Tumors: Distinction of Tumor from Treatment Effect & Fractional Tumor Burden Kathleen M Schmainda1, Melissa A Prah1, Mona Al-Gizawiy1, Jennifer M. Connelly2,3, Wade M Mueller3, Michael J. Schmainda4, Todd Jensen5, Timothy Dondlinger4. 1Radiology, 2Neurology, 3Neurosurgery, Medical College of Wisconsin (MCW), Milwaukee, Wisconsin; 4Imaging Biometrics (IB) LLC, Elm Grove, WI, 5Jensen Informatics, Brookfield WI. Introduction: The overall goal of this U01 project (CA176110) is to develop and validate both standard and novel perfusion-weighted MRI (PWI) and diffusion-weighted MRI (DWI) biomarkers for evaluation of brain tumors and their response to therapies. Highlights from this past year include (1) publication of a manuscript describing the theory, feasibility and advantages of SPICE (spiral perfusion imaging with consecutive echoes) technology over standard DSC-MRI perfusion methods1 and (2) spatially-correlated tissue results validating the ability of rCBV (relative cerebral blood volume) to distinguish glioblastoma (GBM) from treatment effect, thereby enabling the estimation of (3) fractional tumor burden (FTB), a new biomarker2 to predict outcomes. Finally, (4) our industrial partner, Imaging Biometrics LLC (IB), developed IB RadTechTM, a customizable workflow wizard, used to streamline and automate workflows. Methods: Forty-four tissue samples from 15 patients, confirmed as either pure GBM (n=34) or pure treatment effect (TE: n=10), were spatially matched to pre-surgical MRI, which included diffusion and perfusion MRI scans. Biopsy locations were determined via a StealthStation® S7™ surgical navigation unit. Generalized estimating equations and ROC analysis was performed to determine thresholds to distinguish GBM from TE. Using the rCBV thresholds to distinguish GBM from TE, the fraction of tumor within enhancing lesions can be spatially delineated. FTB maps were created for 22 additional patients with GBM with an average of 32 days after completion of radiation therapy. The FTB processing was incorporated as an IB RadTechTM workflow, a tool described in more detail in a separate “QIN Tools” abstract and demonstration. Kaplan Meier survival analysis was performed to determine if FTB, mean rCBV, and/or MGMT methylation status were predictive of overall or profession-free survival (OS, PFS). Results: Pathologic diagnosis identified pure TE in 10 and pure GBM in 34 tissue samples. While rCBV and CBF showed significant differences between TE and GBM (p < 0.01), additional perfusion and diffusion metrics did not (Table 1)3. Based on ROC analysis, thresholds were determined for the creation of FTB maps. As shown (Figure 1), they enable the delineation of the portion of enhancing tumor that is true tumor burden. Figure 2 shows that patients with an FTB > 75% had a worse OS (and PFS, not shown), while MGMT methylation status or mean rCBV (not shown) were not predictive of either OS or PFS4. [1] Paulson ES et al., Tomography 2(4):295-307 (2016). [2] Hu LS et al., Neuro Oncology 14(7):919-930 (2012). [3] Prah MA et al. Proc ISMRM (2015). [4] Prah MA et al., Proc ISMRM (2017) p 707.
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. 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. We are also developing novel acquisition and analysis methods such as vessel size and vessel architecture imaging, advanced diffusion weighted imaging as well multi-parametric and machine learning approaches. Our decision support tools rely on our expertise in image segmentation and registration methods, tumor growth models as well as predictive models of tumor response to therapy. Our “double baseline” studies have established the variability in the DCE-MRI, DSC-MRI and diffusion MRI based parameters (Ktrans, rCBV, rCBF, MTT, ADC, FA) in patients scanned 2-5 days apart. These studies also establish best practices for image analysis to achieve maximal robustness Our novel image acquisition methods include a double-echo DSC and DCE sequences and a multi-shell, multi-directional diffusion sequence that help us better elucidate the tumor microenvironment. In addition to progress in the image acquisition arena, we have also made significant strides in image analysis and informatics. We have also developed a number of open-source image analysis tools for tumor segmentation and registration, multimodal atlases, personalized tumor growth models and hardware and software approaches to improve image resolution. Our radiomics pipeline with documentation and tutorials are publicly available as python packages and Docker containers. We have developed a deep learning framework for segmentation and for radiogenomic analysis. The tools are available as stand-alone modules as well as Slicer modules.
MGH continues to participate actively in all the working groups of the QIN. In conjunction with our ITCR and Leidos funded projects, our challenge platform is supporting a number of QIN challenges. Early Tumour Perfusion and Diffusion Evaluated in Multi-modal Imaging following Radiosurgery for Metastatic Brain Cancer Catherine Coolens1234, Brandon Driscoll4, Warren Foltz12, Noha Sinno1, Xingying Wang1, Igor Svistoun1, Caroline Chung45 1 Radiation Medicine Program, Princess Margaret Cancer Centre, UHN, Toronto, ON 2 Department of Radiation Oncology, University of Toronto, Toronto, ON 3 Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, ON 4 QIPCM Program, Techna Institute, UHN, Toronto, ON 5 Depart. of Radiation Oncology, MD Anderson Cancer Centre, Houston, Texas, USA Introduction: Early change in tumour perfusion and diffusion following stereotactic radiosurgery (SRS) is a potential biomarker of response. However, efforts for quantitative model-based measures of DCE and DWI parameters have shown variable findings to-date that may reflect variability in the MR acquisition and/or analysis method. This work describes the use of a voxel-based, multi-modality GPU architecture to include various complimentary solute transport processes such as perfusion and diffusion into a common framework. This is anticipated to improve accuracy and robustness of the early imaging biomarker predictions. Methods: Patients treated with SRS as part of REB-approved clinical trials underwent volumetric DCE-CT, DCE-MRI and DWI-MRI scans at baseline, then 7 and 20 days post-SRS. As DCE-CT is considered a good standard for tracer-kinetic validation given its signal linearity, we compared 3D pharmaco-kinetic parameter maps using a modified Tofts model (ktrans, ve, Vb, AUC) from both modalities as well as the correlation between apparent diffusion coefficient ADC values from DWI-MRI and the extravascular, extracellular volume (Ve) from DCE imaging. A total of 14 tumours in 9 patients were evaluated. All imaging was co-registered to T1-Gad tumour contours and voxelwise correlations evaluated inside the GTV by Pearson correlation and Bland-Altman comparison. Results: Voxel-wise analysis showed statistically significant correlations in Ktrans (P<0.001) between DCE-CT and DCE-MRI over all imaging time points as well as excellent agreement with very little bias. Statistically significant correlations were also present between ADC/Ve, MRI but a large variation was present across tumors (R2: 0.15-0.8) and disappeared altogether when reviewing the mean ADC only hence disregarding tumor heterogeneity. Conclusion: Utility of a common analysis platform has shown statistically higher correlations between pharmaco-kinetic parameters than has previously been reported but is highlighting the need for a better understanding of the tumor microenvironment to improve biomarker sensitivity. Predicting Response of Low Grade Gliomas to Therapy from MRI Images using Convolutional Neural Networks (CNNs) Zeynettin Akkus, Jiří Sedlář, Panagiotis Korfiatis, Bradley J. Erickson Mayo Clinic, Rochester, MN, USA Introduction: Previous studies have shown that 1p/19q co-deletion is a strong prognostic and predictive molecular marker for positive tumor response to chemotherapy as well as radiotherapy in low-grade gliomas (LGGs). Therefore, prediction of 1p/19q status is crucial for effective treatment planning of LGGs. Presently, determining the 1p/19q status requires surgical biopsy followed by histopathologic analysis. In this study, we provide an alternative, non-invasive method to predict 1p/19q status of LGGs from images of MRI using CNNs. Materials and Methods: The study consists of 75 LGG patients with 3 slices each (n=225 slices), who have biopsy proven 1p/19q status, 48 non-deleted (n/n) and 27 co-deleted (d/d), and pre-operative post-contrast T1 and T2 weighted MRI images. First, we registered post contrast T1 (T1C) images to T2 images. Second, we segmented tumors from 3 consecutive 2D slices that contains the largest tumor size and centered each slice in a standard bounding box size (125x175 pixels) to maintain consistency. Additionally, a binary morphological dilation was applied to the segmentations to include tumor boundaries. The dataset was divided into training (n=3x40 slices), validation (n=20% of training), and test (n=3x14) sets. The training data was balanced for equal probability (20 n/n, 20 d/d) of classes and then augmented with 20 iterations of random translational shift, rotation, and horizontal and vertical flips to increase the training samples (n=2520). Moreover, we shuffle the training data to counter over fitting and provide generalization in each epoch (iterations over all examples). Finally, we trained a multi-branch CNN architecture until 95% accuracy in training and validation with at least 100 epochs and evaluated its performance on the test set. We also compared the performance of our method to the performance of a classical machine-learning algorithm using support vector machine (SVM) classifier with greedy feature selection. Using seven selected features (from intensity-based features, local binary patterns, Gabor filters, Laplacian of Gaussian, gray-level co-occurrence matrix, and boundary sharpness) the SVM classifier was trained and tested on the same data.
Conclusion: CNNs, which rely on hierarchical feature extractions from raw data, offer promising results on prediction of the 1p/19q status of LGGs based on pre-operative diagnostic MRI images. As seen in the results deep learning algorithm performs better than the classical machine-learning algorithm using SVM. Identifying 1p/19q status of LGG non-invasively from MRI images would allow selecting the efficient treatment for each patient and avoiding the need for surgical biopsy. Determination of MGMT Methylation status of GBMs from MRI using machine learning Panagiotis Korfiatis and Bradley Erickson Mayo Clinic, Rochester MN Intro: In glioblastoma multiforme (GBM) patients, methylation of the O6 -methylguanine methyltransferase (MGMT) gene promoter is positively correlated with an increased effectiveness of current standard of care and a better prognosis. This is typically determined based on tissue studies, but fails in up to 10% of cases due to inadequate tissue. In this study, we compare a classic convolutional neural network (VGG-19) versus a more sophisticated deep neural network application (ResNet) for predicting MGMT status from unedited/unsegmented MRI. Methods: A total of 155 T2-weighted fast spin-echo MRI examinations were utilized in this study (66 methylated and 89 unmethylated tumors). The slices corresponding to the enhancing part of the tumor were identified and selected. Subsequently, the data were N4 bias corrected 2. All the data were normalized to the median value of each scan. A VGG-Net architecture as well as a residual deep convolutional neural network were implemented. Residual deep network architectures enable deeper networks to effectively deal with the vanishing gradient problem3. A 50-layer network as well as a network with 21 layers were implemented. In both architectures the tanh activation function was used as the activation classifier, while a sigmoid was used at the output layer. A kernel size of 3x3 was utilized. Both networks were run for 100 epochs for each fold. All the slices containing enhancing tumor were used. The dataset was randomly split into 2 parts: One part was used to train and validate the classifier (120 patients, 2567 slices MGMT methylated and 1969 non MGMT methylated), and one for testing (35 patients, 663 slices MGMT methylated and 473 non MGMT methylated). Results: A fivefold cross validation was used to evaluate the accuracy of the proposed algorithm. Accuracy, precision, recall and the area under the ROC curve were utilized as the proposed metrics evaluated on the test dataset. During the testing phase an Az of 0.91% ± 0.04% and 0.95% ± 0.02% was achieved for the shallow and deeper architectures, respectively. The VGG network performed worse that the ResNet architecture achieve an Az value of 0.79% ± 0.06%. Table 1 summarizes the results on the test dataset for both architectures considered. Conclusion: Results demonstrate that deep neural networks can predict MGMT methylation status in preoperative GBM tumors, thus providing a new noninvasive imaging biomarker useful for management of patients. Furthermore, Deep Learning appears to perform better than traditional machine learning, and ResNets appear to substantially improve performance for a given number of epochs or layers.
Table 1: Results for the two ResNet and the VGG architectures considered in this study. Multi-Imaging Risk Biomarkers in Poor Prognostic Head and Neck Cancer Feifei Teng, Jae Lee, Madhava Aryal, Choonik Lee, Michelle Mierza, Matthew Schipper, Avraham Eisbruch, Yue Cao Department of Radiation Oncology, University of Michigan Introduction: 18F-deoxyglucose (FDG) PET, dynamic contrast enhanced (DCE) MRI and diffusion weighted (DW) MRI each identify unique imaging risk factors for treatment outcomes in head-and-neck cancer (HNC), such as high FDG uptake, low blood volume (LBV), and low apparent diffusion coefficient (LADC). However, most recently completed or ongoing clinical trials in HNC rely on a single imaging modality to define targets for local intensified radiation treatment (boosting). Few studies have investigated whether the image risk biomarkers are independent predictors for treatment outcomes and have any spatial correspondence. We hypothesize that a single imaging modality could not adequately identify the risk subvolumes for failure in poor prognostic head and neck cancers. Purpose: This study aimed to investigate the spatial correspondence of imaging risk factors (high FDG uptake, LBV and LADC in HNC and their early response to chemoradiation therapy (CRT), and to determine the implication of this overlap or lack thereof for individualized adaptive CRT. Materials and methods: Forty patients with poor prognostic HNC (long-history smokers, and/or HPV-) were enrolled in an IRB-approved randomized phase II clinical trial. FDG-PET/CT scans were performed pre-RT, and DCE and DW MRI were acquired pre-RT and after 2 weeks of CRT. All image volumes were co-registered to the post-Gd T1-weighted (T1-Gd) image volume. Gross tumor volume (GTV) was drawn on the T1-Gd images. FDG metabolic tumor sub-volume (MTV) was defined as the standard uptake value (SUV) > 50% of the maximum SUV. Low blood tumor volume (LBV) was defined as BV<7.6 ml/100g within GTV. Low ADC tumor volume (LADCV) was defined as ADC>0.1x10-6 and <1200x10-6 mm2/s within GTV. Spatial overlap between each pair of MTV, LBV and LADCV were evaluated using the Dice similarity coefficient (DC). Percentages of the sub-volumes of imaging risk factors within the GTV were calculated. Responses of the imaging metrics to 2 weeks of CRT were evaluated. Voxel-by-voxel correlations in each tumor volume between each pair of the image parameters were also analyzed. Results: Prior to RT, the MTV, LBV and LADCV of primary tumor had respective median values of 15.5, 10.9 and 27.3 cc, which were only 22%, 16% and 40% of the median GTV (69.1 cc), respectively. DCs between MTV and LBV, between MTV and LADCV, and between LBV and LADCV were 10%, 45%, and 15% in the primary tumors, respectively, indicating low spatial correspondence between the imaging risk factors. The union of MTV, LBV, and LADCV was approximately 77% of GTV, and the union of LBV and LADCV was 56% of GTV, suggesting boosting the union of 2 or 3 imaging risk sub-volumes pre-RT could increase toxicity. However, after 2 weeks of CRT, while the GTV was reduced by 20%, the LBV, LADCV and union of the two were reduced by 47%, 61% and 50%, respectively, consistent with ~50% of MTV reduction in a similar time interval reported previously. Assessing the imaging risk factors at week 2 during CRT could identify the resistant sub-volume of tumor for individualized intensified treatment. Also, there was a great variation of voxel-level correlation between each of pair of image parameters of interested between individual tumors. The median values of voxel-by-voxel cross correlation coefficients (CCC) in individual primary GTVs were 0.11 (from -0.14 to 0.55) between FDG-SUV and BV, and -0.37 (from -0.61 to 0.16) between FDG-SUV and ADC, suggesting inter-tumoral heterogeneity. A single imaging modality may not be able to capture all the risk sub-volumes due to tumor heterogeneity. Conclusion: High FDG uptake, low BV and low ADC as imaging risk biomarkers of HNC identify largely distinct sub-volumes of HNC. Radiation boosting the union of the 3 imaging risk sub-volumes pre-RT could increase toxicity. One proposed solution may be to adaptively boost the union at week-2 during RT, generating a volume representing the persisting high-risk for failure and which is reduced by ~50% compared to pre-RT. Further analysis will be carried out to compare the image metrics with patterns of failure. Quantitative Radiomics of Breast MRI: image analysis and machine learning for imaging-genomic association discovery studies and precision medicine Maryellen Giger, Hui Li, Karen Drukker, Li Lan, Yuan Ji, Yitan Zhu, Suzanne Conzen, and the TCGA Breast Phenotype Group The University of Chicago, Chicago, IL Introduction: In this poster, we present a summary of our work conducted within the Breast Phenotype Group of The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA) in the NCI. In our studies, we investigated relationships between computer-extracted quantitative MRI radiomic tumor features and various clinical, molecular, and genomic markers of prognosis and risk of recurrence, including gene expression profiles. Adapting the Precision Medicine Initiative into imaging research includes studies in both discovery and translation in order to enable the conversion of current radiological interpretation from that of the “average patient” to the precise interpretation and patient-care management decisions specific to the individual. The goal is to individually detect disease, and then give the right person the right treatment at the right time. Methods: We pursued a two-stage approach of “Discovery” followed by “Translation” in our multi-disciplinary imaging-genomics effort involving radiologists, medical physicists, oncologists, computer scientists, engineers, and computational geneticists. Using the collected de-identified datasets of invasive breast carcinomas from TCGA and TCIA, we investigated relationships between computer-extracted quantitative radiomic MRI features and various clinical, molecular, and genomics markers of prognosis and risk of recurrence, including gene expression profiles. At the time of analysis, 91 biopsy-proven invasive breast cancers from the TCGA had Dynamic Contrast Enhanced DCE-MR images. On these cases, we assessed the predictive ability of the quantitative radiomic MRI features relative to four tasks: (i) pathologic stage, (ii) cancer subtypes, (iii) risk of cancer recurrence, and (iv) genomics. The quantitative radiomics features were automatically extracted from dynamic-contrast-enhanced MR images (DCE-MRI) using methods and algorithms developed at The University of Chicago that automatically segment the tumor from the surrounding parenchymal background within the DCE-MR images and extract lesion characteristics in six phenotypic categories: (i) size (measuring tumor dimensions), (ii) shape (quantifying the 3D geometry), (iii) morphology (margin characteristics), (iv) enhancement texture (describing the heterogeneity within the texture of the contrast uptake in the tumor on the first post-contrast MRIs), (v) kinetic curve assessment (describing the shape of the kinetic curve and assessing the physiologic process of the uptake and washout of the contrast agent in the tumor during the dynamic imaging series, and (vi) enhancement-variance kinetics (characterizing the time course of the spatial variance of the enhancement within the tumor. Results: Tumor size was found to be the most powerful predictor of pathologic stage, but radiomic features that captured biologic behavior also emerged as predictive (e.g., stage I and II vs. stage III yielded an AUC of 0.83). Even after controlling for tumor size, statistically significant trends were observed between enhancement texture (entropy) and molecular subtypes (normal-like, luminal A, luminal B, HER2-enriched, basal-like). Also, use of radiomics in the task of distinguishing between high and low likelihoods of cancer recurrence yielded AUC values of 0.88, 0.76, and 0.68 for MammaPrint, Oncotype DX, and PAM50 risk of relapse based on subtype, respectively. In addition, associations between miRNA expressions and radiomic phenotypes were highly specific in that miRNA expressions were associated with primarily two types of radiomic phenotypes – tumor size and enhancement texture; suggesting that miRNAs may mediate the growth of tumor and the heterogeneity of angiogenesis in tumor. Conclusion: Through an extensive investigation, we identified statistically significant associations between quantitative MRI radiomic features and various clinical, molecular, and genomic features in breast invasive carcinoma. Among the novel findings from these “virtual digital biopsies”, we discovered some highly specific imaging-genomic associations, which may be potentially useful in imaging-based diagnoses that can inform the genetic progress of tumor, in discovery of genetic mechanisms that regulate the development of tumor phenotypes, and, ultimately, in predictive models for clinical use. Quantitative Image Analysis and Deep Learning in Breast Cancer Diagnosis and Response to Therapy Maryellen Giger, Hui Li, Karen Drukker, Natasha Antroprova, Ben Huynh, Kayla Mendel, Hiroyuki Abe, Suzanne Conzen, Li Lan, Alexandra Edwards, John Papaioannou The University of Chicago, Chicago, IL Introduction: The goal of our research is to develop quantitative image-based surrogate markers of breast cancer tumors for use in predicting response to therapy and ultimately aiding in patient management. There is a large variation in the clinical presentation of breast cancer in women, and it has been shown that in many instances, biological characteristics, i.e., features, of the primary tumor correlate with outcome. Methods to assess such biological features for the prediction of outcome, however, may be invasive, expensive or not widely available. Our hypothesis is that MRI-based features obtained through quantitative image analysis will prove useful as non-invasive biomarkers for the assessment of, and prediction of, the response of breast cancer to neoadjuvant therapy. Methods & Results: We modified our diagnostic quantitative MRI analysis software to automatically and objectively calculate pre-, during-, and post-treatment breast cancer tumor characteristics (features) including volumetric, morphological, textural, and kinetic features. We also continued to investigate the robustness of our computer-extracted MRI lesion phenotypes. Our recent robustness study focused on the robustness of features across MR scanners of two different manufacturers, GE (N = 91 cases) and Philips (N = 332 cases), in the prognostic task of distinguishing positive and negative lymph node status and receptor statuses of breast cancers. Our results demonstrated that robustness in values and in performance across MR scanners varies for different features. Additionally, we demonstrated that a classification model trained on a dataset of one MR manufacturer did not always generalize to a dataset of another MR manufacturer, thus requiring further optimization and harmonization. In addition, we participated in the QIN BMMR Challenge, which related MRI-based features (phenotypes) with “risk of recurrence” using a UCSF dataset for training and the I-SPY 1 dataset of 162 cases for testing. Our analysis only used both our automatic, computerized lesion-segmentation algorithms and our feature-extraction algorithms on the breast MRIs. Participating in the Challenge yielded many “lessons learned”, which will be described in the group’s future publication. One of our features – one that automatically assessed the tumor’s most-enhancing volume did well in the I-SPY 1 prediction model. Because of the varied differences between the training and testing datasets, robust merged models were difficult to train. A major component of our research is to continue to develop, evaluate, and compare quantitative breast imaging phenotypes (radiomics) of breast cancers, including segmentation-based features of tumor size, shape, morphology, enhancement texture, and kinetics as well as deep learning features extracted from multiple layers of pre-trained convolutional neural networks (CNN). We investigated CNN features extracted with pre-trained CNNs of AlexNet and VGGNet in the task of breast malignancy assessment and response assessment. We found that the combination of the segmentation features and the CNN features enhanced the performance in breast MRI malignancy assessment resulting in an AUC value of 0.91 (se=0.01), a statistically significant improvement over the performance of the CNN method alone. In estimating likelihood of response to therapy, with a leave-one-out cross validation (by subject), we found that the classifier trained on VGGNet features from strictly the pre-contrast time point performed the best, with an AUC of 0.85 (SD = 0.033). The remaining classifiers resulted in AUCs ranging from 0.71 (SD = 0.028) to 0.82 (SD = 0.027). Overall, we found the pre-contrast time point to be the most effective at predicting response to therapy and that including additional contrast time points moderately reduces variance. Conclusion: Continued investigation of the robustness of lesion-based and CNN-based features is warranted. Findings indicate that pre-trained convolutional neural networks can be used to extract image characteristics from breast DCE-MR images relevant to both diagnostic and response assessment, and that combining the intuitive segmentation-based features with CNN features enhances the computerized decision making. Quantitative Imaging for Assessing Breast Cancer Response to Treatment NCI Quantitative Imaging Network Nola Hylton (PI), David Newitt, Ella Jones, Lisa Wilmes, Wen Li, Laura Esserman University of California, San Francisco Yüklə 4,95 Mb. Dostları ilə paylaş:
|
