Imaging Informatics
Year in Review 2016


William Hsu, PhD

        University of California, Los Angeles, CA 


Charles E. Kahn, Jr., MD, MS

        University of Pennsylvania, Philadelphia, PA

in collaboration with

Table of Contents


The authors thank Bradley J. Erickson, MD, PhD, Curtis P. Langlotz, MD, PhD, and Liron Pantanowitz, MD for their suggestions.

Interoperability and Communication

Health information exchange associated with improved emergency department care through faster accessing of patient information from outside organizations

Everson J, et al.  J Am Med Inform Assoc 2016 Aug 12. pii: ocw116.

Each 1-hour reduction in access time reduced ED visit length by 53 minutes, decreased the likelihood of imaging or admission, and reduced average charges by $1187. Faster access to information from outside organizations through HIE improved care and reduced utilization in the ED.

Standardized reporting in IR: a prospective multi-institutional pilot study   

McWilliams JP, et al.  J Vasc Interv Radiol. 2016 Sep 23. doi: 10.1016/j.jvir.2016.07.016.

Standardized report adoption rates increased when reports were simplified by reducing the number of forced fill-in fields. Referring physicians preferred the standardized reports, whereas interventional radiologists preferred standard narrative reports.

Patient access to online radiology reports: frequency and sociodemographic characteristics associated with use

Miles RC, et al. Acad Radiol 2016 Sep;23(9):1162-9.

51% of patients with radiology reports viewed a report online.  Women, patients 25–39 years old, and English speakers were most likely to do so. Higher viewing rates were associated with viewing other types of reports and lower rates were associated with characteristics of traditionally underserved patient populations.

"Chasing a ghost": factors that influence primary care physicians to follow up on incidental imaging findings 

Zafar HM, et al.  Radiology 2016 May 17:152188.  doi: 10.1148/radiol.2016152188

PCPs cited factors that influence how they communicate and manage incidental imaging findings. Radiologists can help PCPs and patients to best use the information in imaging reports.

International telepathology consultation: Three years of experience between the University of Pittsburgh Medical Center and KingMed Diagnostics in China 

Zhao C, et al. J Pathol Inform 2015 Nov 27;6:63

1561 cases were submitted for telepathology consultation from China in 2012-2014; 61% were referred by pathologists, 37% by clinicians, and 2% by patients.  Final diagnoses rendered by U.S. pathologists were identical in 26% of cases and significantly modified (treatment plan altered) in 51% of cases.  

Exploring virtual reality technology and the Oculus Rift for the examination of digital pathology slides

Farahani N, et al.  J Pathol Inform. 2016 May 4;7:22

Digital slides from whole slide imaging (WSI) platforms typically are viewed in 2-D using desktop monitors or mobile devices.  Using a virtual reality (VR) headset to view and navigate pathology whole slide images is feasible, but image resolution was limited.  

Natural Language Processing

Using automatically extracted information from mammography reports for decision-support

Bozkurt S, et al. J Biomed Inform 2016; 62:224-31.

Evaluation of an automated system for extracting BI-RADS descriptors from mammography reports that are then used as inputs into a Bayesian network to predict probability of breast cancer. Automated system achieved high concordance (0.95) when compared to model predictions using manually identified BI-RADS descriptors and a high accuracy rate for predicting BI-RADS category.

A natural language processing tool for large-scale data extraction from echocardiography reports

Nath et al, PLoS One 2016; 11(4):e0153749

Development of a natural language processing system that extracts 59 quantitative and 21 qualitative data elements from a large cohort of 15,116 echocardiography reports from 1,684 patients. Despite the large variety of data elements, the system achieved a recall of 92-99.9% percent and a precision of >97%.

Radiomics and Radiogenomics

Reproducibility of radiomics for deciphering tumor phenotype with imaging

Zhao B, et al.  Sci Rep 2016; 6:23428.

Radiomic features are highly dependent on acquisition techniques and parameters, which influence reproducibility. Unique, same-day repeat computed tomography dataset was used to evaluate the agreement of a common set of 89 radiomic features. Radiomic features were reproducible over a wide range of CT imaging parameters. Imaging settings of thinner slice thickness and smoother reconstruction yielded more consistent results across radiomic features.

Integration of bioinformatics and imaging informatics for identifying rare PSEN1 variants in Alzheimer's disease
Nho K et al, 2016; 9 Suppl 1:30.

Study involved 757 non-Hispanic Caucasian participants who provided peripheral blood that was used for whole genome sequencing and underwent a high resolution T1-weighted structural MRI. Single variant analysis showed that PSEN1 rare variants collectively show a significant association with brain atrophy in regions preferntially affected by late-onset Alzheimer’s disease.

Machine Learning

Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases.  

Janowczyk A, Madabhushi A.  J Pathol Inform 2016 Jul 26;7:29.

Review paper describes the current state-of-the-art of using deep learning to address analysis challenges in digital pathology. Compared to handcrafted features that need to be manually defined from the imaging data, deep learning provides a domain agnostic approach to combine feature extraction and classification. An open source framework (Caffe) was used to generate different network architectures and evaluated in a variety of use cases: nuclei segmentation; epithelium segmentation; tubule segmentation; lymphocyte detection; mitosis detection; invasive ductal carcinoma detection; and lymphoma classification.

How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? 

Cho J, et al.  arXiv: 1511.06348

Access to large medical imaging datasets is often difficult, raising the question how much data is needed to train a deep learning model to achieve necessary high accuracy? A convolutional neural network (CNN) using six different sizes of training data set (5, 10, 20, 50, 100, and 200) was generated and then evaluated in its ability to classify the body part depicted in 6000 CT images. A learning curve approach was used to represent classification performance as a function of sample size.

Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning.  

Shin HC, et al.  IEEE Trans Med Imaging. 2016 May;35(5):1285-98.

The authors evaluated different convolutional neural networks (CNN) architectures with models that contained 5,000 to 160 million parameters for two specific computer-aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification.

Computer-Aided Detection and Diagnosis

Progress in fully automated abdominal CT interpretation.  

Summers RM.  AJR Am J Roentgenol 2016 Jul;207(1):67-79.

Fully automated assessment of organs, lymph nodes, adipose tissue, muscle, bowel, spine, and tumors in abdominal CT has progressed rapidly. Automated analysis ultimately leads to fully automated interpretation.

Automated detection, localization, and classification of traumatic vertebral body fractures in the thoracic and lumbar spine at CT.  

Burns JE, et al.  Radiology 2016 Jan;278(1):64-73.

A fully automated system detected and anatomically localized vertebral body fractures in the thoracic and lumbar spine on CT images. The sensitivity for fracture localization to the correct vertebra was 0.92 with a false-positive rate of 1.6.

Best Papers

Radiology:  2015 Alexander Margulis Award

The Alzheimer structural connectome: changes in cortical network topology with increased amyloid plaque burden.

Prescott JW, et al. for the Alzheimer’s Disease Neuroimaging Initiative.  Radiology 2014 Oct;273(1):175-84.

Increased amyloid burden, as measured with florbetapir PET imaging, is related to changes in the topology of the large-scale cortical network architecture of the brain, as measured with graph theoretical metrics of DTI tractography, even in the preclinical stages of Alzheimer disease.

SIIM 2016:  Roger A Bauman, MD Award

Proving value in radiology: experience developing and implementing a shareable open source registry platform

Ginchoya J, et al.  

Registries are playing an increasingly important role in tracking information such as radiation dosages and providing evidence to prove value on the imaging services provided. OpenMRS was used as ihe basis of a Y90 treatment registry that captures data elements collected at initial encounter, treatment, and follow-up and standardized using RadLex and SNOMED. A pilot database of 20 patients permitted the study of trends in liver enzymes to identify the occurrence of liver failure after Y90 treatment.

Journal of Digital Imaging:  Best Paper of the Year

Retrospective review of the drop in observer detection performance over time in lesion-enriched experimental studies

Taylor-Phillips S, et al.  J Digit Imaging 2015 Feb;28(1):32-40.  

The vigilance decrement describes a decrease in sensitivity or increase in specificity with time in repetitive visual tasks.  Time per case decreased 9-23% as the reading session progressed (p < 0.005). Sensitivity decreased or specificity increased over the course of reading 100 chest x-rays (p = 0.005), 60 bone fracture x-rays (p = 0.03), and 100 chest CT scans (p < 0.0001). This effect was not found in the shorter mammography sessions with 27 or 50 cases.

SIIM 2016 Hackathon:  Grand Prize

Integrated Patient-Centric Portal

Kamel P.

This 2016 SIIM Hackathon project aims to streamline radiology workflow by integrating data from PACS, electronic medical records, and workflow systems into single unified platform that emphasizes patient-driven workflow.

SPIE 2016

[Conference Finalist for PACS and Imaging Informatics]

A handheld computer-aided diagnosis system and simulated analysis [9789-23]

Student Author: Ming Jian Su, Guangxi Univ. (China)

This computer aided diagnosis system runs on a cell phone and is designed to diagnose skin diseases. Each image is associated with a table of features (e.g., Discrete Cosine Transform, circularity, textural features) extracted from the image. The system off-loads pattern recognition tasks to a Hadoop cluster and explores the ability to pool and learn from previously seen cases to improve diagnostic accuracy.

[Conference Finalists for Computer-Aided Diagnosis track]

Deep convolutional neural networks for automatic coronary calcium scoring in a screening study with low-dose chest CT [9785-36]

Student Author: Nikolas Lessmann, Univ. Medical Ctr. Utrecht (Netherlands)

A novel application of coronary calcium scoring using convolutional neural networks. A large 1,028 chest computed tomography dataset was utilized to train and test the model. A 2.5 dimensional image (three orthogonal 50 mm x 50 mm patches) are classified by three concurrent CNNs. High sensitivity (97.2%) was achieved.

Radiomics versus physician assessment for the early prediction of local cancer recurrence after stereotactic radiotherapy for lung cancer [9785-50]

Student Author:  Sarah Mattonen, Western Univ. (Canada) and Baines Imaging Research Lab., London Regional Cancer Program (Canada)

Application of radiomics to predict early post-stereotactic ablative radiotherapy. A total of 104 radiomic features were considered: 44 were extracted from the peri-consolidative regions and the remaining from the consolidative regions. A study involving 45 cases demonstrated that radiomics features could detect early changes associated with local recurrence that were not considered by physician reviewers.

Other Articles of Note

Imaging informatics: 25 years of progress.  

Agrawal JP, Erickson BJ, Kahn CE Jr.    Yearb Med Inform 2016 Jun 30;Suppl 1:S23-31.

A selective overview of key developments in medical imaging informatics.

Assessing strength of evidence of appropriate use criteria for diagnostic imaging examinations.  

Lacson R, et al. J Am Med Inform Assoc 2016 May;23(3):649-53.

For health information technology tools to fully inform evidence-based decisions, recommendations must be reliably assessed for quality and strength of evidence. A novel annotation framework was effective for grading the strength of evidence supporting appropriate use criteria for diagnostic imaging exams.

A data-driven approach for quality assessment of radiologic interpretations 

Hsu W, et al.  J Am Med Inform Assoc.

The authors present a data-driven approach to automate quality assessment of radiologic interpretations using other clinical information (e.g., pathology) as a reference standard for individual radiologists, subspecialty sections, imaging modalities, and entire departments.

Does integrating nonurgent, clinically significant radiology alerts within the electronic health record impact closed-loop communication and follow-up? 

O'Connor SD, et al.   J Am Med Inform Assoc.

Integrating critical result management software -- Alert Notification of Critical Results (ANCR) -- with an electronic health record (EHR) provides an additional workflow for acknowledging nonurgent, clinically significant results without significant change in rates of closed-loop communication or follow-up of alerts.

Harvard Library of Evidence.

To help clinicians choose the most appropriate imaging test for each patient, Harvard Medical School is launching a publicly accessible digital repository of medical evidence.

Predicting the future — big data, machine learning, and clinical medicine.  

Obermeyer Z, Emanuel EJ.  N Engl J Med 2016 Sep 29;375(13):1216-9.

The algorithms of machine learning, which can sift through vast numbers of variables looking for combinations that reliably predict outcomes, will improve prognosis, displace much of the work of radiologists and anatomical pathologists, and improve diagnostic accuracy.