Deep Learning to Classify Radiology Free-Text Reports.

Purpose To evaluate the performance of a deep learning convolutional neural network (CNN) model compared with a traditional natural language processing (NLP) model in extracting pulmonary embolism (PE) findings from thoracic computed tomography (CT) reports from two institutions. Materials and Methods Contrast material-enhanced CT examinations of the chest performed between January 1, 1998, and January 1, 2016, were selected. Annotations by two human radiologists were made for three categories: the presence, chronicity, and location of PE. Classification of performance of a CNN model with an unsupervised learning algorithm for obtaining vector representations of words was compared with the open-source application PeFinder. Sensitivity, specificity, accuracy, and F1 scores for both the CNN model and PeFinder in the internal and external validation sets were determined. Results The CNN model demonstrated an accuracy of 99% and an area under the curve value of 0.97. For internal validation report data, the CNN model had a statistically significant larger F1 score (0.938) than did PeFinder (0.867) when classifying findings as either PE positive or PE negative, but no significant difference in sensitivity, specificity, or accuracy was found. For external validation report data, no statistical difference between the performance of the CNN model and PeFinder was found. Conclusion A deep learning CNN model can classify radiology free-text reports with accuracy equivalent to or beyond that of an existing traditional NLP model. © RSNA, 2017 Online supplemental material is available for this article.

Automated generation of directed graphs from vascular segmentations.

Automated feature extraction from medical images is an important task in imaging informatics. We describe a graph-based technique for automatically identifying vascular substructures within a vascular tree segmentation. We illustrate our technique using vascular segmentations from computed tomography pulmonary angiography images. The segmentations were acquired in a semi-automated fashion using existing segmentation tools. A 3D parallel thinning algorithm was used to generate the vascular skeleton and then graph-based techniques were used to transform the skeleton to a directed graph with bifurcations and endpoints as nodes in the graph. Machine-learning classifiers were used to automatically prune false vascular structures from the directed graph. Semantic labeling of portions of the graph with pulmonary anatomy (pulmonary trunk and left and right pulmonary arteries) was achieved with high accuracy (percent correct⩾0.97). Least-squares cubic splines of the centerline paths between nodes were computed and were used to extract morphological features of the vascular tree. The graphs were used to automatically obtain diameter measurements that had high correlation (r⩾0.77) with manual measurements made from the same arteries.

A Narrative Review of Sociodemographic Disparities in Relation to PEHR Access.

Personal electronic health records (PEHRs) enable patients access to their own medical records. Differences in access and use of PEHRs may create health disparities. We conducted a narrative literature review regarding the effects of race, language preference, education, income, and homelessness on PEHR usage as well as PEHRs content, particularly stigmatizing language. Of 3177 citations found, 75 articles were relevant. Patient race, language, income, and education predicted PEHR use, which could potentially exacerbate health disparities.

Patient Access to Health Data: A Review of Philosophic and Healthcare Issues.

In 2012 Australia created a national Personal Controlled Electronic Health Record (PCEHR) known as "My Health Record" (MHR). However, MHR has seen low patient utilization. Debate regarding MHR has centered on utility and moral issues (e.g. data privacy). We conducted a narrative review to assess patient perception and clinical utility of PCEHRs worldwide. Results show patient and clinician support for PCEHRs but little evidence of improved outcomes and patient concerns regarding data providence.

Analyzing the Spread of Informatics with PubMed.

We analyzed PubMed citations since 1988 to explore the dissemination of medical/health informatics concepts between countries and across medical domains. We extracted countries from the PubMed author affiliation field to identify and analyze the top 10 informatics publishing countries. We found that the informatics publications are becoming more similar over time and that the rate of exchange across countries has increased with the introduction of e-publishing. Nonetheless, with the exception of machine learning, the impact of core informatics concepts on mainstream medicine and radiology publications remains small.

Designing an Interprofessional Online Course to Foster Learning Health Systems.

The Learning Health Systems (LHS) framework demonstrates the potential for iterative interrogation of health data in real time and implementation of insights into practice. Yet, the lack of appropriately skilled workforce results in an inability to leverage existing data to design innovative solutions. We developed a tailored professional development program to foster a skilled workforce. The short course is wholly online, for interdisciplinary professionals working in the digital health arena. To transform healthcare systems, the workforce needs an understanding of LHS principles, data driven approaches, and the need for diversly skilled learning communities that can tackle these complex problems together.

Document-level classification of CT pulmonary angiography reports based on an extension of the ConText algorithm.

In this paper we describe an application called peFinder for document-level classification of CT pulmonary angiography reports. peFinder is based on a generalized version of the ConText algorithm, a simple text processing algorithm for identifying features in clinical report documents. peFinder was used to answer questions about the disease state (pulmonary emboli present or absent), the certainty state of the diagnosis (uncertainty present or absent), the temporal state of an identified pulmonary embolus (acute or chronic), and the technical quality state of the exam (diagnostic or not diagnostic). Gold standard answers for each question were determined from the consensus classifications of three human annotators. peFinder results were compared to naive Bayes' classifiers using unigrams and bigrams. The sensitivities (and positive predictive values) for peFinder were 0.98(0.83), 0.86(0.96), 0.94(0.93), and 0.60(0.90) for disease state, quality state, certainty state, and temporal state respectively, compared to 0.68(0.77), 0.67(0.87), 0.62(0.82), and 0.04(0.25) for the naive Bayes' classifier using unigrams, and 0.75(0.79), 0.52(0.69), 0.59(0.84), and 0.04(0.25) for the naive Bayes' classifier using bigrams.

Compartment and hub definitions tune metabolic networks for metabolomic interpretations.

BACKGROUND: Metabolic networks represent all chemical reactions that occur between molecular metabolites in an organism's cells. They offer biological context in which to integrate, analyze, and interpret omic measurements, but their large scale and extensive connectivity present unique challenges. While it is practical to simplify these networks by placing constraints on compartments and hubs, it is unclear how these simplifications alter the structure of metabolic networks and the interpretation of metabolomic experiments. RESULTS: We curated and adapted the latest systemic model of human metabolism and developed customizable tools to define metabolic networks with and without compartmentalization in subcellular organelles and with or without inclusion of prolific metabolite hubs. Compartmentalization made networks larger, less dense, and more modular, whereas hubs made networks larger, more dense, and less modular. When present, these hubs also dominated shortest paths in the network, yet their exclusion exposed the subtler prominence of other metabolites that are typically more relevant to metabolomic experiments. We applied the non-compartmental network without metabolite hubs in a retrospective, exploratory analysis of metabolomic measurements from 5 studies on human tissues. Network clusters identified individual reactions that might experience differential regulation between experimental conditions, several of which were not apparent in the original publications. CONCLUSIONS: Exclusion of specific metabolite hubs exposes modularity in both compartmental and non-compartmental metabolic networks, improving detection of relevant clusters in omic measurements. Better computational detection of metabolic network clusters in large data sets has potential to identify differential regulation of individual genes, transcripts, and proteins.

The International Neuromodulation Registry: An Informatics Framework Supporting Cohort Discovery and Analysis.

BACKGROUND: Neuromodulation therapies, such as deep brain stimulation (DBS), spinal cord stimulation (SCS), responsive neurostimulation (RNS), transcranial magnetic stimulation (TMS), transcranial direct stimulation (tDCS), and vagus nerve stimulation (VNS) are used to treat neurological and psychiatric conditions for patients who have failed to benefit from other treatment approaches. Although generally effective, seemingly similar cases often have very different levels of effectiveness. While there is ongoing interest in developing predictors, it can be difficult to aggregate the necessary data from limited cohorts of patients at individual treatment centers. OBJECTIVE: In order to increase the predictive power in neuromodulation studies, we created an informatics platform called the International Neuromodulation Registry (INR). The INR platform has a data flow process that will allow researchers to pool data across multiple centers to enable population health research. METHODS: This custom informatics platform has a Neo4j graph database and includes a harmonization process that allows data from different studies to be aggregated and compared. Users of the INR can download deidentified patient imaging, patient demographic data, device settings, and medical rating scales. The INR supports complex network analysis and patient timeline visualization. RESULTS: The INR currently houses and allows visualization of deidentified imaging and clinical data from hundreds of patients with a wide range of diagnoses and neuromodulation therapies. CONCLUSION: Ultimately, we believe that widespread adoption of the INR platform will improve population health research in neuromodulation therapy.

Feasibility of Natural Language Processing-Assisted Auditing of Critical Findings in Chest Radiology.

OBJECTIVE: Time-sensitive communication of critical imaging findings like pneumothorax or pulmonary embolism to referring physicians is essential for patient safety. The definitive communication is the radiology free-text report. Quality assurance initiatives require that institutions audit these communications, a time-intensive manual task. We propose using a rule-based natural language processing system to improve the process for auditing critical findings communications. METHODS: We present a pilot assessment of the feasibility of using an automated critical finding identification system to assist quality assurance teams' evaluation of critical findings communication compliance. Our assessment is based on chest imaging reports. Critical findings are identified in radiology reports using pyConTextNLP, an open source Python implementation of the ConText algorithm. RESULTS: In our test set, there were 75 reports with critical findings and 591 reports without critical findings. pyConTextNLP correctly identified 69 of the positive cases with 8 false-positives for a sensitivity of 0.92 and a specificity of 0.99. DISCUSSION: Natural language processing can provide valuable assistance to auditing critical findings communications.

Extracting a stroke phenotype risk factor from Veteran Health Administration clinical reports: an information content analysis.

BACKGROUND: In the United States, 795,000 people suffer strokes each year; 10-15 % of these strokes can be attributed to stenosis caused by plaque in the carotid artery, a major stroke phenotype risk factor. Studies comparing treatments for the management of asymptomatic carotid stenosis are challenging for at least two reasons: 1) administrative billing codes (i.e., Current Procedural Terminology (CPT) codes) that identify carotid images do not denote which neurovascular arteries are affected and 2) the majority of the image reports are negative for carotid stenosis. Studies that rely on manual chart abstraction can be labor-intensive, expensive, and time-consuming. Natural Language Processing (NLP) can expedite the process of manual chart abstraction by automatically filtering reports with no/insignificant carotid stenosis findings and flagging reports with significant carotid stenosis findings; thus, potentially reducing effort, costs, and time. METHODS: In this pilot study, we conducted an information content analysis of carotid stenosis mentions in terms of their report location (Sections), report formats (structures) and linguistic descriptions (expressions) from Veteran Health Administration free-text reports. We assessed an NLP algorithm, pyConText's, ability to discern reports with significant carotid stenosis findings from reports with no/insignificant carotid stenosis findings given these three document composition factors for two report types: radiology (RAD) and text integration utility (TIU) notes. RESULTS: We observed that most carotid mentions are recorded in prose using categorical expressions, within the Findings and Impression sections for RAD reports and within neither of these designated sections for TIU notes. For RAD reports, pyConText performed with high sensitivity (88 %), specificity (84 %), and negative predictive value (95 %) and reasonable positive predictive value (70 %). For TIU notes, pyConText performed with high specificity (87 %) and negative predictive value (92 %), reasonable sensitivity (73 %), and moderate positive predictive value (58 %). pyConText performed with the highest sensitivity processing the full report rather than the Findings or Impressions independently. CONCLUSION: We conclude that pyConText can reduce chart review efforts by filtering reports with no/insignificant carotid stenosis findings and flagging reports with significant carotid stenosis findings from the Veteran Health Administration electronic health record, and hence has utility for expediting a comparative effectiveness study of treatment strategies for stroke prevention.

Automatic detection of three-dimensional vascular tree centerlines and bifurcations in high-resolution magnetic resonance angiography.

OBJECTIVES: We sought to develop a simple and robust algorithm capable of automatically detecting centerlines and bifurcations of a three-dimensional (3D) vascular bed. MATERIALS AND METHODS: After necessary preprocessing, an appropriate cost function is computed for all vessel voxels and Dijkstra's minimum-cost-path algorithm is implemented. By back tracing all the minimum-cost paths, centerlines and bifurcation are detected. The detected paths are then split into segments between adjacent nodes (bifurcations or vessel end-points) and smoothed by curve fitting. RESULTS: Application of the algorithm to both simulated 3D vessels and 3D magnetic resonance angiography (MRA) images of an actual intracranial arterial tree produced well-centered vessel skeletons. Quantitative assessment of the algorithm was performed. For the simulated data, the root mean square error for centerline detection is about half a voxel. For the human intracranial MRA data, the sensitivity, positive predictive value (PPV), and accuracy of bifurcation detection were calculated for different cost functions. The best case gave a sensitivity of 91.4%, a PPV of 91.4%, and an RMS error of 1.7 voxels. CONCLUSIONS: To the extent that imperfections are eliminated from the segmented image, the algorithm is effective and robust in automatic and accurate detection of centerlines and bifurcations. The cost function and algorithm used are demonstrated to be an improvement over similar algorithms in the literature.

3D multi-scale vessel enhancement filtering based on curvature measurements: application to time-of-flight MRA.

In this paper we evaluate the use of voxel intensity curvature measurements to enhance vessels in 3D MRA images. We compare a multi-scale discrete kernel filter (MaxCurve) to the Hessian matrix based filter proposed by Frangi and co-workers. The MaxCurve filter is based on the maximum difference between the negative curvature computed along orthogonal lines defined by a 3x3x3 kernel. Filter performance is assessed using measures of vessel and background separation (contrast and the area under the ROC curve). Filter parameters are optimized using a training set of four typical time-of-flight MRA images and tested on a separate set of ten MRA images with the same acquisition parameters. The filters tended to provide good MIP image contrast enhancement. The filters are applied to MRA images acquired with different parameters and field strengths indicating potential usefulness for a variety of images. Overall the discrete kernel and Hessian matrix filter performed quite similarly.

Registration of MR and CT images of the liver: comparison of voxel similarity and surface based registration algorithms.

The purpose of this work was to determine the feasibility and efficacy of retrospective registration of MR and CT images of the liver. The open-source ITK Insight Software package developed by the National Library of Medicine (USA) contains a multi-resolution, voxel-similarity-based registration algorithm which we selected as our baseline registration method. For comparison we implemented a multi-scale surface fitting technique based on the head-and-hat algorithm. Registration accuracy was assessed using the mean displacement of automatically selected point landmarks. The ITK voxel-similarity-based registration algorithm performed better than the surface-based approach with mean misregistration in the range of 7.7-8.4 mm for CT-CT registration, 8.2 mm for MR-MR registration, and 14.0-18.9 mm for MR-CT registration compared to mean misregistration from the surface-based technique in the range of 9.6-11.1 mm for CT-CT registration, 9.2-12.4 mm for MR-MR registration, and 15.2-19.0 mm for MR-CT registration.

Creating a text classifier to detect radiology reports describing mediastinal findings associated with inhalational anthrax and other disorders.

OBJECTIVE: The aim of this study was to create a classifier for automatic detection of chest radiograph reports consistent with the mediastinal findings of inhalational anthrax. DESIGN: The authors used the Identify Patient Sets (IPS) system to create a key word classifier for detecting reports describing mediastinal findings consistent with anthrax and compared their performances on a test set of 79,032 chest radiograph reports. MEASUREMENTS: Area under the ROC curve was the main outcome measure of the IPS classifier. Sensitivity and specificity of an initial IPS model were calculated based on an existing key word search and were compared against a Boolean version of the IPS classifier. RESULTS: The IPS classifier received an area under the ROC curve of 0.677 (90% CI = 0.628 to 0.772) with a specificity of 0.99 and maximum sensitivity of 0.35. The initial IPS model attained a specificity of 1.0 and a sensitivity of 0.04. CONCLUSION: The IPS system is a useful tool for helping domain experts create a statistical key word classifier for textual reports that is a potentially useful component in surveillance of radiographic findings suspicious for anthrax.

Automated lung segmentation in X-ray computed tomography: development and evaluation of a heuristic threshold-based scheme.

RATIONALE AND OBJECTIVES: To develop and evaluate a reliable, fully-automated lung segmentation scheme for application in X-ray computed tomography. MATERIALS AND METHODS: The automated scheme was heuristically developed using a slice-based, pixel-value threshold and two sets of classification rules. Features used in the rules include size, circularity, and location. The segmentation scheme operates slice-by-slice and performs three key operations: (1) image preprocessing to remove background pixels, (2) computation and application of a pixel-value threshold to identify lung tissue, and (3) refinement of the initial segmented regions to prune incorrectly detected airways and separate fused right and left lungs. RESULTS: The performance of the automated segmentation scheme was evaluated using 101 computed tomography cases (91 thick slice, 10 thin slice scans). The 91 thick cases were pre- and post-surgery from 50 patients and were not independent. The automated scheme successfully segmented 94.0% of the 2,969 thick slice images and 97.6% of the 1,161 thin slice images. The mean difference of the total lung volumes calculated by the automated scheme and functional residual capacity plus 60% inspiratory capacity was -24.7 +/- 508.1 mL. The mean differences of the total lung volumes calculated by the automated scheme and an established, commonly used semi-automated scheme were 95.2 +/- 52.5 mL and -27.7 +/- 66.9 mL for the thick and thin slice cases, respectively. CONCLUSION: This simple, fully-automated lung segmentation scheme provides an objective tool to facilitate lung segmentation from computed tomography scans.

Intracranial vessel segmentation from time-of-flight MRA using pre-processing of the MIP Z-buffer: accuracy of the ZBS algorithm.

We evaluate the accuracy of a vascular segmentation algorithm which uses continuity in the maximum intensity projection (MIP) depth Z-buffer as a pre-processing step to generate a list of 3D seed points for further segmentation. We refer to the algorithm as Z-buffer segmentation (ZBS). The pre-processing of the MIP Z-buffer is based on smoothness measured using the minimum chi-square value of a least square fit. Points in the Z-buffer with chi-square values below a selected threshold are used as seed points for 3D region growing. The ZBS algorithm couples spatial continuity information with intensity information to create a simple yet accurate segmentation algorithm. We examine the dependence of the segmentation on various parameters of the algorithm. Performance is assessed in terms of the inclusion/exclusion of vessel/background voxels in the segmentation of intracranial time-of-flight MRA images. The evaluation is based on 490,256 voxels from 14 patients which were classified by an observer. ZBS performance was compared to simple thresholding and to segmentation based on vessel enhancement filtering. The ZBS segmentation was only weakly dependent on the parameters of the initial MIP image generation, indicating the robustness of this approach. Region growing based on Z-buffer generated seeds was advantageous compared to simple thresholding. The ZBS algorithm provided segmentation accuracies similar to that obtained with the vessel enhancement filter. The ZBS performance was notably better than the filter based segmentation for aneurysms where the assumptions of the filter were violated. As currently implemented the algorithm slightly under-segments the intracranial vasculature.

Identifying respiratory findings in emergency department reports for biosurveillance using MetaMap.

Clinical conditions described in patients' dictated reports are necessary for automated detection of patients with respiratory illnesses such as inhalational anthrax and pneumonia. We applied MetaMap to emergency department reports to extract a set of 71 clinical conditions relevant to detection of a lower respiratory outbreak. We indexed UMLS terms in emergency department reports with MetaMap, filtered the indexed output with a specialized lexicon of UMLS terms for the domain, and mapped the clinical conditions of interest to concepts in the lexicon. We compared MetaMap's ability to accurately identify the conditions against a physician's manual annotations and evaluated incorrectly indexed features to determine what additional processing is necessary. MetaMap identified the clinical conditions with a recall of 0.72 and a precision of 0.56. Necessary processing beyond MetaMap's indexing includes finding validation, temporal discrimination, anatomic location discrimination, finding-disease discrimination, and contextual inference. Successful identification of clinical conditions in an emergency department report with MetaMap requires processing techniques specific to the clinical question of interest.

Cue-based assertion classification for Swedish clinical text--developing a lexicon for pyConTextSwe.

OBJECTIVE: The ability of a cue-based system to accurately assert whether a disorder is affirmed, negated, or uncertain is dependent, in part, on its cue lexicon. In this paper, we continue our study of porting an assertion system (pyConTextNLP) from English to Swedish (pyConTextSwe) by creating an optimized assertion lexicon for clinical Swedish. METHODS AND MATERIAL: We integrated cues from four external lexicons, along with generated inflections and combinations. We used subsets of a clinical corpus in Swedish. We applied four assertion classes (definite existence, probable existence, probable negated existence and definite negated existence) and two binary classes (existence yes/no and uncertainty yes/no) to pyConTextSwe. We compared pyConTextSwe's performance with and without the added cues on a development set, and improved the lexicon further after an error analysis. On a separate evaluation set, we calculated the system's final performance. RESULTS: Following integration steps, we added 454 cues to pyConTextSwe. The optimized lexicon developed after an error analysis resulted in statistically significant improvements on the development set (83% F-score, overall). The system's final F-scores on an evaluation set were 81% (overall). For the individual assertion classes, F-score results were 88% (definite existence), 81% (probable existence), 55% (probable negated existence), and 63% (definite negated existence). For the binary classifications existence yes/no and uncertainty yes/no, final system performance was 97%/87% and 78%/86% F-score, respectively. CONCLUSIONS: We have successfully ported pyConTextNLP to Swedish (pyConTextSwe). We have created an extensive and useful assertion lexicon for Swedish clinical text, which could form a valuable resource for similar studies, and which is publicly available.