Search

PurposeTo assess changes in retinopathy of prematurity (ROP) diagnosis in single and serial retinal images.DesignCohort study.ParticipantsCases of ROP recruited from the Imaging and Informatics in Retinopathy of Prematurity (i-ROP) consortium evaluated by 7 graders.MethodsSeven ophthalmologists reviewed both single and 3 consecutive serial retinal images from 15 cases with ROP, and severity was assigned as plus, preplus, or none. Imaging data were acquired during routine ROP screening from 2011 to 2015, and a reference standard diagnosis was established for each image. A secondary analysis was performed using the i-ROP deep learning system to assign a vascular severity score (VSS) to each image, ranging from 1 to 9, with 9 being the most severe disease. This score has been previously demonstrated to correlate with the International Classification of ROP. Mean plus disease severity was calculated by averaging 14 labels per image in serial and single images to decrease noise.Main outcome measuresGrading severity of ROP as defined by plus, preplus, or no ROP.ResultsAssessment of serial retinal images changed the grading severity for > 50% of the graders, although there was wide variability. Cohen's kappa ranged from 0.29 to 1.0, which showed a wide range of agreement from slight to perfect by each grader. Changes in the grading of serial retinal images were noted more commonly in cases of preplus disease. The mean severity in cases with a diagnosis of plus disease and no disease did not change between single and serial images. The ROP VSS demonstrated good correlation with the range of expert classifications of plus disease and overall agreement with the mode class (P= 0.001). The VSS correlated with mean plus disease severity by expert diagnosis (correlation coefficient, 0.89). The more aggressive graders tended to be influenced by serial images to increase the severity of their grading. The VSS also demonstrated agreement with disease progression across serial images, which progressed to preplus and plus disease.ConclusionsClinicians demonstrated variability in ROP diagnosis when presented with both single and serial images. The use of deep learning as a quantitative assessment of plus disease has the potential to standardize ROP diagnosis and treatment.

Persistent avascular retina (PAR) in prematurely born individuals may be a risk factor for late sequelae of retinopathy of prematurity (ROP), including retinal detachment in older childhood and adulthood. Although PAR has been associated with use of vascular endothelial growth factor antagonist therapy for treatment-requiring ROP, the prevalence of this finding in patients without prior ROP treatment is unknown. We performed a cross-sectional study to determine the prevalence of PAR in a cohort of patients 4-8 years of age who were screened for ROP in the neonatal intensive care unit and did not receive treatment. Patients were recruited from an existing population-based cohort and underwent ultra-widefield fluorescein angiography (UWFFA). UWFFA images of 43 eyes of 23 patients were evaluated. Average age at time of evaluation was 6.2 years. PAR was observed in 21 patients (91%). Thirteen eyes (30%) had PAR in zone II; 23 (53%), in zone III. Six eyes (14%) had abnormal vessels without clear PAR. These findings indicate a high prevalence of PAR in patients with a history of ROP screening without treatment.

Purpose To present an overview of current machine learning methods and their use in medical research, focusing on select machine learning techniques, best practices, and deep learning. Methods A systematic literature search in PubMed was performed for articles pertinent to the topic of artificial intelligence methods used in medicine with an emphasis on ophthalmology. Results A review of machine learning and deep learning methodology for the audience without an extensive technical computer programming background. Conclusions Artificial intelligence has a promising future in medicine; however, many challenges remain. Translational Relevance The aim of this review article is to provide the nontechnical readers a layman's explanation of the machine learning methods being used in medicine today. The goal is to provide the reader a better understanding of the potential and challenges of artificial intelligence within the field of medicine.

ObjectiveTo evaluate informatics-enabled quality improvement (QI) strategies for promoting time spent on face-to-face communication between ophthalmologists and patients.MethodsThis prospective study involved deploying QI strategies during implementation of an enterprise-wide vendor electronic health record (EHR) in an outpatient academic ophthalmology department. Strategies included developing single sign-on capabilities, activating mobile- and tablet-based applications, EHR personalization training, creating novel workflows for team-based orders, and promoting problem-based charting to reduce documentation burden. Timing data were collected during 648 outpatient encounters. Outcomes included total time spent by the attending ophthalmologist on the patient, time spent on documentation, time spent on examination, and time spent talking with the patient. Metrics related to documentation efficiency, use of personalization features, use of team-based orders, and note length were also measured from the EHR efficiency portal and compared with averages for ophthalmologists nationwide using the same EHR.ResultsTime spent on exclusive face-to-face communication with patients initially decreased with EHR implementation (2.9 to 2.3 minutes, p = 0.005) but returned to the paper baseline by 6 months (2.8 minutes, p = 0.99). Observed participants outperformed national averages of ophthalmologists using the same vendor system on documentation time per appointment, number of customized note templates, number of customized order lists, utilization of team-based orders, note length, and time spent after-hours on EHR use.ConclusionInformatics-enabled QI interventions can promote patient-centeredness and face-to-face communication in high-volume outpatient ophthalmology encounters. By employing an array of interventions, time spent exclusively talking with the patient returned to levels equivalent to paper charts by 6 months after EHR implementation. This was achieved without requiring EHR redesign, use of scribes, or excessive after-hours work. Documentation efficiency can be achieved using interventions promoting personalization and team-based workflows. Given their efficacy in preserving face-to-face physician-patient interactions, these strategies may help alleviate risk of physician burnout.

Presentation at the 2019 National Library of Medicine Informatics Training Conference on Automated Fundus Image Quality Assessment in Retinopathy of Prematurity.

We assess the ability of a Random Forest model to predict image quality from clinical data alone (i.e., without evaluating images).

Generative adversarial networks (GANs) are a class of unsupervised machine learning algorithms that can produce realistic images from randomly-sampled vectors in a multi-dimensional space. Until recently, it was not possible to generate realistic high-resolution images using GANs, which has limited their applicability to medical images that contain biomarkers only detectable at native resolution. Progressive growing of GANs is an approach wherein an image generator is trained to initially synthesize low resolution synthetic images (8x8 pixels), which are then fed to a discriminator that distinguishes these synthetic images from real downsampled images. Additional convolutional layers are then iteratively introduced to produce images at twice the previous resolution until the desired resolution is reached. In this work, we demonstrate that this approach can produce realistic medical images in two different domains; fundus photographs exhibiting vascular pathology associated with retinopathy of prematurity (ROP), and multi-modal magnetic resonance images of glioma. We also show that fine-grained details associated with pathology, such as retinal vessels or tumor heterogeneity, can be preserved and enhanced by including segmentation maps as additional channels. We envisage several applications of the approach, including image augmentation and unsupervised classification of pathology.

The purpose of this study was to implement an algorithm for automatically assessing the quality of fundus images, and to evaluate its performance on a set of Retinopathy of Prematurity (ROP) images compared to expert assessment. The computer-based algorithm assessed image quality with strong correlation to the expert consensus rank, although not as high as individual expert correlation with the consensus rank.

The differentiation factor NEL-like molecule-1 (NELL-1) has been reported as osteoinductive in multiple in vivo preclinical models. Bone morphogenetic protein (BMP)-2 is used clinically for skeletal repair, but in vivo administration can induce abnormal, adipose-filled, poor-quality bone. We demonstrate that NELL-1 combined with BMP2 significantly optimizes osteogenesis in a rodent femoral segmental defect model by minimizing the formation of BMP2-induced adipose-filled cystlike bone. In vitro studies using the mouse bone marrow stromal cell line M2-10B4 and human primary bone marrow stromal cells have confirmed that NELL-1 enhances BMP2-induced osteogenesis and inhibits BMP2-induced adipogenesis. Importantly, the ability of NELL-1 to direct BMP2-treated cells toward osteogenesis and away from adipogenesis requires intact canonical Wnt signaling. Overall, these studies establish the feasibility of combining NELL-1 with BMP2 to improve clinical bone regeneration and provide mechanistic insight into canonical Wnt pathway activity during NELL-1 and BMP2 osteogenesis. The novel abilities of NELL-1 to stimulate Wnt signaling and to repress adipogenesis may highlight new treatment approaches for bone loss in osteoporosis.

This file contains Figures S1â S7 as well as Tables S2â S9. (PDF 18206 kb)

Average precision of cell detection over a variety of experimental conditions. Table S2. Average precision of cell detection over the whole worm gonadal arms. Table S3. Average overlap of cell segmentations across different experimental conditions. Table S4. Benchmarking EdU quantification accuracy. Table S5. Performance comparisons with other software. Table S6. Performance scored following ground truths provided by different users. Table S7. Number of connected components of DNA image in G1-, S-, G2-, and M-phase cells. Table S8. Pairwise comparisons of number of connected components in DNA images. Table S9. DNA spatial extent in G1-, S-, G2-, and M phase cells. Table S10. Pairwise comparisons of DNA spatial extent in G1-, S-, G2-, and M-phase cells. Table S11. Sensitivity and specificity of classifier. (PDF 114Â kb)

Benchmarking of cell row counter accuracy for C. elegans gonadal arms. (A) Overlay of microscope image (white signal, derived from the DNA stain DAPI) with segmentations color-coded by cell row position from the distal end (left), as computed automatically using our counter. (B) Size of the MZ scored manually vs size computed through the automatic counter. A small amount of noise (0.5 cell rows) was added in order to aid visualization of overlapping data. Positions scored manually and automatically are in close agreement; average percent deviation is 9.4Â %. Diagonal shown for reference in red. (PDF 152Â kb)

Average precision of cell detection over a variety of experimental conditions. Table S2. Average precision of cell detection over the whole worm gonadal arms. Table S3. Average overlap of cell segmentations across different experimental conditions. Table S4. Benchmarking EdU quantification accuracy. Table S5. Performance comparisons with other software. Table S6. Performance scored following ground truths provided by different users. Table S7. Number of connected components of DNA image in G1-, S-, G2-, and M-phase cells. Table S8. Pairwise comparisons of number of connected components in DNA images. Table S9. DNA spatial extent in G1-, S-, G2-, and M phase cells. Table S10. Pairwise comparisons of DNA spatial extent in G1-, S-, G2-, and M-phase cells. Table S11. Sensitivity and specificity of classifier. (PDF 114Â kb)

Parismi GUI screenshot. (PDF 1381Â kb)

Cell detection via HOG features. (A) HOG weights learned from germ cells. The left two panels display the component of the weight vector learned from xy sections, while the right two panels correspond to xz sections. (B) Variation of training samples. The left panel corresponds to the 20 highest-scoring training samples and the right panel corresponds to the 20 lowest-scoring training samples. (C) Eight curated MZs are sufficient for accurate training of the cell detector. (PDF 697Â kb)

Comparison of frozen or vibratome section segmentations. (A) Active contour segmentation of frozen sections is often inaccurate (red arrows) due to poor cell separation in the original image. (B) Vibratome sections enable clearer staining and thus more accurate segmentation (yellow arrows). (PDF 749Â kb)

Examples of varying DNA morphologies across C. elegans germ cells. Cells shown were taken at different phases of mitosis (A) or meiosis (B). (PDF 334Â kb)

Benchmarking of cell row counter accuracy for C. elegans gonadal arms. (A) Overlay of microscope image (white signal, derived from the DNA stain DAPI) with segmentations color-coded by cell row position from the distal end (left), as computed automatically using our counter. (B) Size of the MZ scored manually vs size computed through the automatic counter. A small amount of noise (0.5 cell rows) was added in order to aid visualization of overlapping data. Positions scored manually and automatically are in close agreement; average percent deviation is 9.4Â %. Diagonal shown for reference in red. (PDF 152Â kb)

Comparison of hand-constructed segmentation produced independently by three users. The segmentations were performed in three dimensions; the same representative slice is shown for the three users. (PDF 23Â kb)

Cell detection via HOG features. (A) HOG weights learned from germ cells. The left two panels display the component of the weight vector learned from xy sections, while the right two panels correspond to xz sections. (B) Variation of training samples. The left panel corresponds to the 20 highest-scoring training samples and the right panel corresponds to the 20 lowest-scoring training samples. (C) Eight curated MZs are sufficient for accurate training of the cell detector. (PDF 697Â kb)

Example output of various programs on C. elegans gonadal arm. (A) Ilastik after training with ~230 brush strokes, showing original DNA signal in white, and pixels classified as foreground (respectively background) in green (respectively red). Arrow points to nuclei that do not appear separated by background pixels, and arrowhead to a set of pixels that are classified as background instead of foreground. (B-E) The same representative slice from original DNA signal (B), hand-constructed segmentation (C), Parismi segmentation (D), and MINS segmentation (E), in which segments tend to say confined to the interior of the nuclear domain delimited by DNA signal. (PDF 420Â kb)

Precision-recall curves benchmarking cell center detection. (A-B) Precision-recall trade off as the detection threshold is varied for C. elegans MZ cells (A) and mouse embryo (B). Based on these curves, an automatic detection threshold can be chosen that yields high precision (e.g. P = 0.98); automatic detection using this threshold can be followed by manual curation using Parismi’s annotation tool to add cells missed by the detector. Red circles show recall and precision for each of the three manual segmentations performed on the same set of cells (see main text). (PDF 117 kb)

Example output of various programs on C. elegans gonadal arm. (A) Ilastik after training with ~230 brush strokes, showing original DNA signal in white, and pixels classified as foreground (respectively background) in green (respectively red). Arrow points to nuclei that do not appear separated by background pixels, and arrowhead to a set of pixels that are classified as background instead of foreground. (B-E) The same representative slice from original DNA signal (B), hand-constructed segmentation (C), Parismi segmentation (D), and MINS segmentation (E), in which segments tend to say confined to the interior of the nuclear domain delimited by DNA signal. (PDF 420Â kb)

Precision-recall curves benchmarking cell center detection. (A-B) Precision-recall trade off as the detection threshold is varied for C. elegans MZ cells (A) and mouse embryo (B). Based on these curves, an automatic detection threshold can be chosen that yields high precision (e.g. P = 0.98); automatic detection using this threshold can be followed by manual curation using Parismi’s annotation tool to add cells missed by the detector. Red circles show recall and precision for each of the three manual segmentations performed on the same set of cells (see main text). (PDF 117 kb)

Comparison of hand-constructed segmentation produced independently by three users. The segmentations were performed in three dimensions; the same representative slice is shown for the three users. (PDF 23Â kb)

Global blindness exacts an enormous financial and social cost on developing countries. Reducing the prevalence of blindness globally requires a set of strategies that are different from those typically used in developed countries. This was the subject of the 2013 Knapp symposium at the American Ophthalmological Society Annual Meeting, and this article summarizes the presentations of epidemiologists, health care planners, and ophthalmologists. It explores a range of successful strategies from the multinational Vision 2020 Initiative to disease-specific schemes in cataract, trachoma control, infectious corneal ulceration, cytomegalovirus retinitis, and retinopathy of prematurity. In each example, the importance of an attitudinal change set toward public health becomes clear. There is reason for optimism in the struggle against global blindness in large measure because of innovative programs such as those described here.

BackgroundMobile phones have great potential for medical education, as they allow health care providers and students to access resources efficiently at the precise time at the point-of-care to help in informed decision making. ObjectiveThe objective of the study was to evaluate the prevalence of mobile phone usage among medical residents and to explore their attitudes, perceptions, and the challenges they experience when using mobile phones in academic and clinical practice. MethodsA cross-sectional survey was conducted on all 133 residents in 17 different specialties across two large academic hospitals in Riyadh, Saudi Arabia. The Web-based validated questionnaire measured mobile phone platform preferences, and their uses in general and medical practice. The perception of confidentiality and safety impact of using mobile phones for communication and accessing patient’s data was also explored, alongside challenges of use and how residents learn to use their mobile phone. ResultsWith a response rate of 101/133 (75.9%) and mean age of 27.8 (SD 3.0) years, we found that 100/101 (99.0%) of participants were mobile phone users with mean duration of use of 5.12 (SD 2.4) years, and a range from 1 to 12 years. There was no significant difference in use between male and female respondents. A negative linear correlation was found between age and use duration (P=.004). The most common operating system used by participants was the iOS platform (55/101, 54.5%), with English the most commonly used language to operate residents’ mobile phones (96/100, 96.0%) despite their native language being Arabic. For communication outside medical practice, chatting applications such as WhatsApp matched phone calls as most commonly used tools (each 88/101, 87.1%). These were also the primary tools for medical communication, but used at a lower rate (each 65/101, 64.4%). In medical practice, drug (83/101, 82.2%) and medical (80/101, 79.2%) references and medical calculation applications (61/101, 60.4%) were the most commonly used. Short battery life (48/92, 52%) was the most common technical difficulty, and distraction at least on a weekly basis (54/92, 58%) was the most likely side effect of using a mobile phone in medical practice. Practically, all participants agreed with the idea of integrating medical staff mobile phones with the hospital information system. Most residents described themselves as self-learners, while half learned from peers, and a quarter learned from the Internet. Only 7/101 (6.9%) had received formal training on the medical use of mobile phones. Over half of residents thought it was safe to discuss patients over their personal, nonencrypted email. ConclusionsMobile phone use among medical residents has become almost universal in academic and clinical settings. Thus, academic and health care institutions should support proper utilization of these devices in medical training and point-of-care decision making, while continuing to protect patient confidentiality.

Electronic health record (EHR) systems have significant potential advantages over traditional paper-based systems, but they require that providers assume responsibility for data entry. One significant barrier to adoption of EHRs is the perception of slowed data-entry by providers. This study compares the speed of data-entry using computer-based templates vs. paper for a large eye clinic, using 10 subjects and 10 simulated clinical scenarios. Dataentry into the EHR was significantly slower (p

Segmentation accuracy for C. elegans germ cells. Overlay of microscope image (white signal, derived from the DNA stain DAPI) with hand-constructed segmentation (green, top row) or automatic segmentation (red, bottom row). Average overlap is 74Â %. 7 z slices are shown, which cover the cell visible in the center of each slice. (PDF 101Â kb)

DNA morphologies of cells classified as G1-, S-, G2-, and M-phase based on DNA and EdU content. (PDF 3296Â kb)

Segmentation accuracy for C. elegans germ cells. Overlay of microscope image (white signal, derived from the DNA stain DAPI) with hand-constructed segmentation (green, top row) or automatic segmentation (red, bottom row). Average overlap is 74Â %. 7 z slices are shown, which cover the cell visible in the center of each slice. (PDF 101Â kb)

DNA morphologies of cells classified as G1-, S-, G2-, and M-phase based on DNA and EdU content. (PDF 3296Â kb)