Search

The medical imaging community has embraced Machine Learning (ML) as evidenced by the rapid increase in the number of ML models being developed, but validating and deploying these models in the clinic remains a challenge. The engineering involved in integrating and assessing the efficacy of ML models within the clinical workflow is complex. This paper presents a general-purpose, end-to-end, clinically integrated ML model deployment and validation system implemented at UCSF. Engineering and usability challenges and results from 3 use cases are presented.A generalized validation system based on free, open-source software was implemented, connecting clinical imaging modalities, the Picture Archiving and Communication System (PACS), and an ML inference server. ML pipelines were implemented in NVIDIA’s Clara Deploy framework with results and clinician feedback stored in a customized XNAT instance, linked within PACS. Prospective clinical validation studies of 3 ML models were conducted, with data routed from multiple clinical imaging modalities and PACS. Completed validation studies provided expert clinical feedback on model performance and usability, plus system reliability and performance metrics.Clinical validation of ML models entails assessing model performance, impact on clinical infrastructure, robustness, and usability. Study results must be easily accessible to participating clinicians but remain outside the clinical record. Building a system that generalizes and scales across multiple ML models takes the concerted effort of software engineers, clinicians, data scientists, and system administrators, and benefits from the use of modular open-source software. The present work provides a template for institutions looking to translate and clinically validate ML models in the clinic, together with required resources and expected challenges.Author summaryAcademic medical centers gather and store vast quantities of digital data, and with the increase in accessibility of Machine Learning (ML) techniques, there has been an explosion of ML model development in the medical imaging community. Most of this work remains in research, though, and connecting ML models to the clinic for testing on live patient data and integration into the clinical workflow remains a challenge and impedes clinical impact. We present a general-purpose system, implemented and deployed at UCSF, for in-clinic validation of ML models and their incorporation into patient care. This work, based on free and open-source software packages, can serve as a template for other institutions looking to solve ML’s “last mile” problem and move their models out of research and into the clinic.

BackgroundDiagnostic classification of diffuse gliomas now requires an assessment of molecular features, often including IDH-mutation and 1p19q-codeletion status. Because genetic testing requires an invasive process, an alternative noninvasive approach is attractive, particularly if resection is not recommended. The goal of this study was to evaluate the effects of training strategy and incorporation of biologically relevant images on predicting genetic subtypes with deep learning.MethodsOur dataset consisted of 384 patients with newly diagnosed gliomas who underwent preoperative MRI with standard anatomical and diffusion-weighted imaging, and 147 patients from an external cohort with anatomical imaging. Using tissue samples acquired during surgery, each glioma was classified into IDH-wildtype (IDHwt), IDH-mutant/1p19q-noncodeleted (IDHmut-intact), and IDH-mutant/1p19q-codeleted (IDHmut-codel) subgroups. After optimizing training parameters, top performing convolutional neural network (CNN) classifiers were trained, validated, and tested using combinations of anatomical and diffusion MRI with either a 3-class or tiered structure. Generalization to an external cohort was assessed using anatomical imaging models.ResultsThe best model used a 3-class CNN containing diffusion-weighted imaging as an input, achieving 85.7% (95% CI: [77.1, 100]) overall test accuracy and correctly classifying 95.2%, 88.9%, 60.0% of the IDHwt, IDHmut-intact, and IDHmut-codel tumors. In general, 3-class models outperformed tiered approaches by 13.5%-17.5%, and models that included diffusion-weighted imaging were 5%-8.8% more accurate than those that used only anatomical imaging.ConclusionTraining a classifier to predict both IDH-mutation and 1p19q-codeletion status outperformed a tiered structure that first predicted IDH-mutation, then 1p19q-codeletion. Including apparent diffusion coefficient (ADC), a surrogate marker of cellularity, more accurately captured differences between subgroups.

BackgroundDifferentiating treatment-induced injury from recurrent high-grade glioma is an ongoing challenge in neuro-oncology, in part due to lesion heterogeneity. This study aimed to determine whether different MR features were relevant for distinguishing recurrent tumor from the effects of treatment in contrast-enhancing lesions (CEL) and non-enhancing lesions (NEL).MethodsThis prospective study analyzed 291 tissue samples (222 recurrent tumor, 69 treatment-effect) with known coordinates on imaging from 139 patients who underwent preoperative 3T MRI and surgery for a suspected recurrence. 8 MR parameter values were tested from perfusion-weighted, diffusion-weighted, and MR spectroscopic imaging at each tissue sample location for association with histopathological outcome using generalized estimating equation models for CEL and NEL tissue samples. Individual cutoff values were evaluated using receiver operating characteristic curve analysis with 5-fold cross-validation.ResultsIn tissue samples obtained from CEL, elevated relative cerebral blood volume (rCBV) was associated with the presence of recurrent tumor pathology (P < 0.03), while increases in normalized choline (nCho) and choline-to-NAA index (CNI) were associated with the presence of recurrent tumor pathology in NEL tissue samples (P < 0.008). A mean CNI cutoff value of 2.7 had the highest performance, resulting in mean sensitivity and specificity of 0.61 and 0.81 for distinguishing treatment-effect from recurrent tumor within the NEL.ConclusionAlthough our results support prior work that underscores the utility of rCBV in distinguishing the effects of treatment from recurrent tumor within the contrast enhancing lesion, we found that metabolic parameters may be better at differentiating recurrent tumor from treatment-related changes in the NEL of high-grade gliomas.

Based on the expanding set of applications for hyperpolarized carbon-13 (HP-13 C) MRI, this work aims to communicate standardized methodology implemented at the University of California, San Francisco, as a primer for conducting reproducible metabolic imaging studies of the prostate and brain. Current state-of-the-art HP-13 C acquisition, data processing/reconstruction and kinetic modeling approaches utilized in patient studies are presented together with the rationale underpinning their usage. Organized around spectroscopic and imaging-based methods, this guide provides an extensible framework for handling a variety of HP-13 C applications, which derives from two examples with dynamic acquisitions: 3D echo-planar spectroscopic imaging of the human prostate and frequency-specific 2D multislice echo-planar imaging of the human brain. Details of sequence-specific parameters and processing techniques contained in these examples should enable investigators to effectively tailor studies around individual-use cases. Given the importance of clinical integration in improving the utility of HP exams, practical aspects of standardizing data formats for reconstruction, analysis and visualization are also addressed alongside open-source software packages that enhance institutional interoperability and validation of methodology. To facilitate the adoption and further development of this methodology, example datasets and analysis pipelines have been made available in the supporting information.

Purpose The purpose of this study was to characterize tissue-specific alterations in metabolism of hyperpolarized (HP) gluconeogenic precursors 13 C-lactate and 13 C-pyruvate by rat liver and kidneys under conditions of fasting or insulin-deprived diabetes. Methods Seven normal rats were studied by MR spectroscopic imaging of both HP 13 C-lactate and 13 C-pyruvate in both normal fed and 24 h fasting states, and seven additional rats were scanned after induction of diabetes by streptozotocin (STZ) with insulin withdrawal. Phosphoenolpyruvate carboxykinase (PEPCK) expression levels were also measured in liver and kidney tissues of the STZ-treated rats. Results Multiple sets of significant signal modulations were detected, with graded intensity in general between fasting and diabetic states. An approximate two-fold reduction in the ratio of 13 C-bicarbonate to total 13 C signal was observed in both organs in fasting. The ratio of HP lactate-to-alanine was markedly altered, ranging from a liver-specific 54% increase in fasting, to increases of 69% and 92% in liver and kidney, respectively, in diabetes. Diabetes resulted in a 40% increase in renal lactate signal. STZ resulted in 5.86-fold and 2.73-fold increases in PEPCK expression in liver and kidney, respectively. Conclusion MRI of HP 13 C gluconeogenic precursors may advance diabetes research by clarifying organ-specific roles in abnormal diabetic metabolism. Magn Reson Med 77:1429-1437, 2017. © 2016 International Society for Magnetic Resonance in Medicine.

Infiltrating low-grade gliomas (LGGs) are a class of terminal central nervous system tumors that comprise malignant neuroglia. Histopathological diagnosis of tumor grade is performed using criteria set by the World Health Organization (WHO) and is based on factors that include nuclear atypia, proliferative capacity, tumor neovascularization, and necrosis.1 The clinical outcome for patients with LGG is variable, with some lesions following a more indolent disease course, while others recur more rapidly and often after undergoing malignant transformation (MT) to a higher grade.2 Lesions that have thus transformed to a grade III anaplastic glioma or grade IV secondary glioblastoma multiforme (GBM) are managed with additional, more aggressive treatments. To date, little is known regarding the nature of recurrent disease, and the most significant prognostic factors for patients diagnosed with LGG are the presence of somatic driver mutations in isocitrate dehydrogenase 1 and 2 (IDH1/2) oncogenes and the codeletion of the 1p and 19q chromosomal arms, which have been associated with increased survival and sensitivity to the treatment given.2–5 IDH mutations have been implicated as an initiating event in gliomagenesis and are key to reprogramming the tumor epigenome and metabolome, largely through neomorphic production and accumulation of 2-hydroxyglutarate.6 Given uncertainties regarding the effectiveness of options available for treating LGGs, IDH mutations have garnered significant attention as a targetable therapeutic pathway, and there are several novel therapies on the horizon.7–9 Until these therapeutics become available, the clinical mainstay of treatment for patients with LGG consists of surgical resection, with radiation therapy and alkylating chemotherapy being typically reserved for recurrences that were either subtotally resected or have undergone MT. MRI is an integral component of brain tumor diagnosis and monitoring. Recent advances in state-of-the-art techniques have enabled the noninvasive assessment of tumor morphology and physiology, with primary GBM having been studied most extensively. Although the ability to noninvasively detect MT would be of significant clinical interest for diagnosis and treatment planning, few studies have focused on addressing this question. Multiparametric MRI holds significant promise for comprehensively characterizing the structural and physiological properties of the tumor. We hypothesize that these advanced imaging techniques may also provide improved characterization of MT in patients with recurrent LGG. The objective of this study was to establish multiparametric MRI profiles of patients with tumors prior to image-guided surgery in order to relate metrics obtained from these methodologies to histopathological grade. Volumetric regions with abnormal imaging features were calculated to provide a robust assessment of the entire recurrent tumor lesion. Regions of viable tumor were targeted for image-guided tissue sampling in order to associate in vivo parameters with histopathological features and to strengthen our understanding of the link between glioma tumor biology and parameters from noninvasive imaging.

BACKGROUND Previous studies of solid tumors have been restricted in their ability to map how heterogeneous cell populations evolved within the tumor in three-dimensional (3D) space due to insufficient sampling, typically one sample per tumor, and a lack of knowledge of where within the tumor the sample was obtained. Knowledge of the extensivity of heterogeneity and how it is spatially distributed is crucial for assessing whether a therapeutic target is truly tumor-wide, and for exploring how mutations relate to heterogeneity in the local microenvironment. METHODS We developed a novel platform to integrate and visualize in 3D multi-omics data generated from each of 8–10 spatially mapped samples per tumor. Together, the 171 samples collected using this approach from 16 adult diffuse glioma at diagnosis and recurrence form a novel resource – the 3D Glioma Atlas. RESULTS By maximally sampling the tumor geography without excluding samples based on low cancer cell fraction (CCF), we identify a subpopulation of glioblastoma with pervasively lower CCF likely excluded by other atlases, such as the TCGA, that used stringent CCF cutoffs. Exome sequencing of 3D-mapped samples from lower-grade tumors revealed that clonal expansions are typically spatially segregated, implying minimal tumor-wide intermixing of genetically heterogenous cells. Heterogeneity is less spatially segregated for faster-growing high-grade tumors, suggesting that cell populations expand in these tumors differently. Recurrent low-grade tumors have greater intratumoral mutational heterogeneity than newly diagnosed tumors, though this did not translate into greater dissimilarity in gene expression profiles for recurrent tumors, suggesting minimal functional impact of this additional mutational diversity on gene expression. CONCLUSIONS The delineation of spatial patterns of heterogeneity that our work provides enables more informed interpretation of biopsies and greater insight into the factors shaping intratumoral variation of gene expression patterns. Ongoing work is exploring the spatial patterning of amplification events and the tumor microenvironment.

Infiltrating gliomas are primary brain tumors that express significant biological and clinical heterogeneity in adults, which complicates their treatment and prognosis. Characterization of tumor subtypes using spectroscopic analysis may assist in predicting malignant transformation and quantification of response to therapy. To implement an automated algorithm for classification of metabolomic profiles for the classification of glioma pathological grades and the prediction of malignant progression using spectra obtained by high-resolution magic angle spinning (HR-MAS) spectroscopy of patient-derived tissue samples. 237 image-guided tissue samples were obtained from 152 patients who underwent surgery for newly diagnosed or recurrent glioma and analyzed via HR-MAS spectroscopy. Orthogonal projection to latent structures discriminant analysis was used as a classifier and the variable-influence-on-projection values were evaluated to identify signature spectral regions. The accuracy of classifiers developed for discriminating glioma subtypes was 68% for newly diagnosed grade II versus III samples; 86 and 92% for new and recurrent grade III versus IV, respectively; 95% for newly diagnosed grade II versus IV; and 88% for recurrent grade II versus IV lesions. Classifiers distinguished between samples from newly diagnosed vs. recurrent lesions with an accuracy of 78% for grade III and 99% for grade IV glioma. Classifying metabolomic profiles for new and recurrent glioma without prior assumptions regarding spectral components identified candidate in vivo biomarkers for use in assessing changes that are likely to impact treatment decisions.

Infiltrating low grade gliomas (LGGs) are heterogeneous in their behavior and the strategies used for clinical management are highly variable. A key factor in clinical decision-making is that patients with mutations in the isocitrate dehydrogenase 1 and 2 (IDH1/2) oncogenes are more likely to have a favorable outcome and be sensitive to treatment. Because of their relatively long overall median survival, more aggressive treatments are typically reserved for patients that have undergone malignant progression (MP) to an anaplastic glioma or secondary glioblastoma (GBM). In the current study, ex vivo metabolic profiles of image-guided tissue samples obtained from patients with newly diagnosed and recurrent LGG were investigated using proton high-resolution magic angle spinning spectroscopy (1H HR-MAS). Distinct spectral profiles were observed for lesions with IDH-mutated genotypes, between astrocytoma and oligodendroglioma histologies, as well as for tumors that had undergone MP. Levels of 2-hydroxyglutarate (2HG) were correlated with increased mitotic activity, axonal disruption, vascular neoplasia, and with several brain metabolites including the choline species, glutamate, glutathione, and GABA. The information obtained in this study may be used to develop strategies for in vivo characterization of infiltrative glioma, in order to improve disease stratification and to assist in monitoring response to therapy.

We present a precision medicine application developed for multiple sclerosis (MS): the MS BioScreen. This new tool addresses the challenges of dynamic management of a complex chronic disease; the interaction of clinicians and patients with such a tool illustrates the extent to which translational digital medicine-that is, the application of information technology to medicine-has the potential to radically transform medical practice. We introduce 3 key evolutionary phases in displaying data to health care providers, patients, and researchers: visualization (accessing data), contextualization (understanding the data), and actionable interpretation (real-time use of the data to assist decision making). Together, these form the stepping stones that are expected to accelerate standardization of data across platforms, promote evidence-based medicine, support shared decision making, and ultimately lead to improved outcomes.

Interpretation of changes in the T1- and T2-weighted MR images from patients with newly diagnosed glioblastoma (GBM) treated with standard of care in conjunction with anti-angiogenic agents is complicated by pseudoprogression and pseudoresponse. The hypothesis being tested in this study was that 3D H-1 magnetic resonance spectroscopic imaging (MRSI) provides estimates of levels of choline, creatine, N-acetylaspartate (NAA), lactate and lipid that change in response to treatment and that metrics describing these characteristics are associated with survival. Thirty-one patients with newly diagnosed GBM and being treated with radiation therapy (RT), temozolomide, erlotinib and bevacizumab were recruited to receive serial MR scans that included 3-D lactate edited MRSI at baseline, mid-RT, post-RT and at specific follow-up time points. The data were processed to provide estimates of metrics representing changes in metabolite levels relative to normal appearing brain. Cox proportional hazards analysis was applied to examine the relationship of these parameters with progression free survival (PFS) and overall survival (OS). There were significant reductions in parameters that describe relative levels of choline to NAA and creatine, indicating that the treatment caused a decrease in tumor cellularity. Changes in the levels of lactate and lipid relative to the NAA from contralateral brain were consistent with vascular normalization. Metabolic parameters from the first serial follow-up scan were associated with PFS and OS, when accounting for age and extent of resection. Integrating metabolic parameters into the assessment of patients with newly diagnosed GBM receiving therapies that include anti-angiogenic agents may be helpful for tracking changes in tumor burden, resolving ambiguities in anatomic images caused by non-specific treatment effects and for predicting outcome.

Quantitative analysis of magnetic resonance spectroscopic imaging (MRSI) data provides maps of metabolic parameters that show promise for improving medical diagnosis and therapeutic monitoring. While anatomical images are routinely reconstructed on the scanner, formatted using the DICOM standard, and interpreted using PACS workstations, this is not the case for MRSI data. The evaluation of MRSI data is made more complex because files are typically encoded with vendor-specific file formats and there is a lack of standardized tools for reconstruction, processing, and visualization. SIVIC is a flexible open-source software framework and application suite that enables a complete scanner-to-PACS workflow for evaluation and interpretation of MRSI data. It supports conversion of vendor-specific formats into the DICOM MR spectroscopy (MRS) standard, provides modular and extensible reconstruction and analysis pipelines, and provides tools to support the unique visualization requirements associated with such data. Workflows are presented which demonstrate the routine use of SIVIC to support the acquisition, analysis, and delivery to PACS of clinical1H MRSI datasets at UCSF.

Background The heterogeneous biology of glioblastoma (GBM) emphasizes the need for imaging methods to assess tumor burden and assist in evaluating individual patients. The purpose of this study was to investigate early changes in metrics from 3D 1H magnetic resonance spectroscopic imaging (MRSI) data, compare them with anatomic lesion volumes, and determine whether they were associated with survival for patients with newly diagnosed GBM receiving a multimodality treatment regimen. Methods Serial MRI and MRSI scans provided estimates of anatomic lesion volumes and levels of choline, creatine, N-acetylaspartate, lactate, and lipid. The association of metrics derived from these data with survival was assessed using Cox proportional hazards models with adjustments for age, Karnofsky performance score, and extent of resection. Temporal changes in parameters were evaluated using a Wilcoxon signed rank test. Results Anatomic lesion volumes at the post-radiotherapy (RT) scan, metabolic lesion volume at mid-RT and post-RT scans, as well as metrics describing levels of choline, lactate, and lipid were associated with overall survival. There was a significant reduction in the enhancing lesion volume, increase in T2 lesion volume from mid-RT to post-RT, and decrease in parameters describing metabolite levels during these early time points. Conclusion The MRSI data provided metrics that described the effects of treatment on the metabolic lesion burden and were associated with overall survival. This suggests that adding these parameters to standard assessments of changes in anatomic lesion volumes could contribute to making early decisions about the efficacy of such combination therapies.

Background:A new method has been used to obtain human sensory evoked-responses whose time-domain waveforms have been undetectable by previous methods. These newly discovered evoked-responses have durations that exceed the time between the stimuli in a continuous stream, thus causing an overlap which, up to now, has prevented their detection. We have named them "A-waves", and added a prefix to show the sensory system from which the responses were obtained (visA-waves, audA-waves, somA-waves).Results:When A-waves were studied as a function of stimulus repetition-rate, it was found that there were systematic differences in waveshape at repetition-rates above and below the psychophysical region in which the sensation of individual stimuli fuse into a continuity. The fusion phenomena is sometimes measured by a "Critical Fusion Frequency", but for this research we can only identify a frequency-region [which we call the STZ (Sensation-Transition Zone)]. Thus, the A-waves above the STZ differed from those below the STZ, as did the sensations.Study of the psychophysical differences in auditory and visual stimuli, as shown in this paper, suggest that different stimulus features are detected, and remembered, at stimulation rates above and below STZ.Conclusion:The results motivate us to speculate that:1) Stimulus repetition-rates above the STZ generate waveforms which underlie "fusion-memory" whereas rates below the STZ show neuronal processing in which "flash-memory" occurs.2) These two memories differ in both duration and mechanism, though they may occur in the same cell groups.3) The differences in neuronal processing may be related to "figure" and "ground" differentiation.We conclude that A-waves provide a novel measure of neural processes that can be detected on the human scalp, and speculate that they may extend clinical applications of evoked response recordings. If A-waves also occur in animals, it is likely that A-waves will provide new methods for comparison of activity of neuronal populations and single cells.