Eye Tracking for Deep Learning Segmentation Using Convolutional Neural Networks.

Deep learning with convolutional neural networks (CNNs) has experienced tremendous growth in multiple healthcare applications and has been shown to have high accuracy in semantic segmentation of medical (e.g., radiology and pathology) images. However, a key barrier in the required training of CNNs is obtaining large-scale and precisely annotated imaging data. We sought to address the lack of annotated data with eye tracking technology. As a proof of principle, our hypothesis was that segmentation masks generated with the help of eye tracking (ET) would be very similar to those rendered by hand annotation (HA). Additionally, our goal was to show that a CNN trained on ET masks would be equivalent to one trained on HA masks, the latter being the current standard approach. Step 1: Screen captures of 19 publicly available radiologic images of assorted structures within various modalities were analyzed. ET and HA masks for all regions of interest (ROIs) were generated from these image datasets. Step 2: Utilizing a similar approach, ET and HA masks for 356 publicly available T1-weighted postcontrast meningioma images were generated. Three hundred six of these image + mask pairs were used to train a CNN with U-net-based architecture. The remaining 50 images were used as the independent test set. Step 1: ET and HA masks for the nonneurological images had an average Dice similarity coefficient (DSC) of 0.86 between each other. Step 2: Meningioma ET and HA masks had an average DSC of 0.85 between each other. After separate training using both approaches, the ET approach performed virtually identically to HA on the test set of 50 images. The former had an area under the curve (AUC) of 0.88, while the latter had AUC of 0.87. ET and HA predictions had trimmed mean DSCs compared to the original HA maps of 0.73 and 0.74, respectively. These trimmed DSCs between ET and HA were found to be statistically equivalent with a p value of 0.015. We have demonstrated that ET can create segmentation masks suitable for deep learning semantic segmentation. Future work will integrate ET to produce masks in a faster, more natural manner that distracts less from typical radiology clinical workflow.

MRI characteristics associated with high-grade myxoid liposarcoma.

AIM: To identify magnetic resonance imaging (MRI) features differentiating high-grade (>5% round-cell component) from low-grade myxoid liposarcomas (LPS) (≤5% round-cell component). MATERIALS AND METHODS: Informed consent was waived. Patients with myxoid LPS and MRI before biopsy, neoadjuvant therapy, and surgery were included retrospectively. High-grade components were recorded from histological specimens by a pathologist (24 years of experience). Images were evaluated by a senior radiologist (>12 years of experience) for tumour size, location, tissue layer, and MRI features (signal intensity, heterogeneity, margin, and perilesional characteristics). Descriptive statistics, Fisher's exact test to identify associations with a round-cell component, and multivariate logistic regression to identify independent predictors of high-grade tumours were used. RESULTS: Thirty-one patients (16 women [mean 51.1 years; range 19-79 years] and 15 men [mean 45.5 years; range 18-95 years]) with myxoid LPS (23 low-grade, eight high-grade) were included. All high-grade lesions had lipid signal, a peritumoural capsule and peritumoural contrast enhancement, and more commonly exhibited heterogeneous signal; however, the average size of ≥10 cm was the strongest independent indicator of high-grade status (odds ratio [OR], 14.6; 95% confidence interval [CI]: 1.6, 131). CONCLUSION: Size ≥10 cm is most strongly associated with high-grade myxoid LPS (round-cell component >5%). Other features possibly differentiating high-grade from low-grade status include lesion margin, lipid signal, and perilesional characteristics.

Bayesian estimation of cerebral perfusion using reduced-contrast-dose dynamic susceptibility contrast perfusion at 3T.

BACKGROUND AND PURPOSE: DSC perfusion has been increasingly used in conjunction with other contrast-enhanced MR applications and therefore there is need for contrast-dose reduction when feasible. The purpose of this study was to establish the feasibility of reduced-contrast-dose brain DSC perfusion by using a probabilistic Bayesian method and to compare the results with the commonly used singular value decomposition technique. MATERIALS AND METHODS: Half-dose (0.05-mmol/kg) and full-dose (0.1-mmol/kg) DSC perfusion studies were prospectively performed in 20 patients (12 men; 34-70 years of age) by using a 3T MR imaging scanner and a gradient-EPI sequence (TR/TE, 1450/22 ms; flip angle, 90°). All DSC scans were processed with block circulant singular value decomposition and Bayesian probabilistic methods. SNR analysis was performed in both half-dose and full-dose groups. The CBF, CBV, and MTT maps from both full-dose and half-dose scans were evaluated qualitatively and quantitatively in both WM and GM on coregistered perfusion maps. Statistical analysis was performed by using a t test, regression, and Bland-Altman analysis. RESULTS: The SNR was significantly (P < .0001) lower in the half-dose group with 32% and 40% reduction in GM and WM, respectively. In the half-dose group, the image-quality scores were significantly higher in Bayesian-derived CBV (P = .02) and MTT (P = .004) maps in comparison with block circulant singular value decomposition. Quantitative values of CBF, CBV, and MTT in Bayesian-processed data were comparable and without a statistically significant difference between the half-dose and full-dose groups. The block circulant singular value decomposition-derived half-dose perfusion values were significantly different from those of the full-dose group both in GM (CBF, P < .001; CBV, P = .02; MTT, P = .02) and WM (CBF, P < .001; CBV, P = .003; MTT, P = .01). CONCLUSIONS: Reduced-contrast-dose (0.05-mmol/kg) DSC perfusion of the brain is feasible at 3T by using the Bayesian probabilistic method with quantitative results comparable with those of the full-dose protocol.

Predictive role of imaging in sentinel lymph node dissection for melanoma.

A retrospective study of 67 patients with metastatic melanoma was performed to evaluate if imaging from lymphoscintigraphy could predict a higher miss rate if only the most radioactive node were removed. Following protocol for sentinel node biopsy, the surgeon resected all lymph nodes containing radioactivity > 10% of the most radioactive node. A correlation was performed between the radioactive counts of the lymph nodes and the presence of metastases. The percentage of cases in which the most radioactive node was negative for metastasis on pathology was calculated. Two nuclear medicine physicians read the images from lymphoscintigraphy specifically to determine if the first lymph node visualized became less intense than other nodes on later images. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated. In 13 of 67 (19%) patients, the most radioactive lymph node was negative for metastasis while a less radioactive node contained metastatic disease. Consensus reading by the nuclear medicine physicians determined that in 9 cases, the first lymph node visualized became less intense than another lymph node on later images. Of the 9 cases, 4 were true positive and 5 were false positive when correlated with intraoperative count rate and pathology. Of the cases where the most radioactive node was not positive on histopathology (n = 13), the consensus reading by the nuclear medicine physicians reported 4 of them (31%). Imaging by lymphoscintigram had a sensitivity 31%, specificity 91%, positive predictive value 44%, and negative predictive value 85% for predicting whether the most radioactive lymph node at surgery would be negative for metastasis at pathology. We conclude that in patients with melanoma, lymphoscintigraphy has high specificity and negative predictive value but modest sensitivity and positive predictive value for detecting when the sentinel node will not be the most radioactive lymph node during sentinel lymph node dissection. These findings support that dynamic imaging by lymphoscintigraphy has a role in surgical planning but that the imaging protocol could benefit from further optimization.

Enhanced visualization processing: effect on workflow.

RATIONALE AND OBJECTIVES: Soft-copy viewing of digital radiographs allows for image processing to improve visualization of anatomy and lesions, but it can take more time than film-based viewing. Enhanced visualization processing (EVP) was developed to increase the latitude of an image without reducing the vital contrast, potentially reducing the need for the radiologist to manipulate images. This study examined the influence of processing radiographic images with EVP on workflow in a picture archiving and communications system (PACS). MATERIALS AND METHODS: Portable computed radiographic chest images were obtained and processed either with EVP or without. A security camera with a videocassette recorder was positioned above the PACS workstation. Four radiologists reviewed the images during their normal work schedule. The current diagnostic image was used to determine if the case contained EVP or non-EVP images. The videotapes of the sessions were reviewed to determine diagnostic viewing times and how zoom and/or window and level manipulation was used. RESULTS: Viewing time was significantly longer for the non-EVP than the EVP cases. The difference occurred with all readers. Window and level manipulation was used on 35% of the EVP and 41% of the non-EVP images. Zoom was used on 64% of the EVP and 69% of the non-EVP images. Average time spent using zoom and window and level manipulation was significantly shorter with the EVP than with the non-EVP images. CONCLUSION: EVP of chest images displayed on PACS monitors significantly improved workflow as measured by viewing time. EVP decreased use of window and level manipulation and zooming and the amount of time each one was used.

Evaluation of a flat CRT monitor for use in radiology.

Medical radiographs based on familiar projection techniques are planar images traditionally displayed by placing on a flat surface viewbox. Presenting these planar images in digital form on a traditional monitor with a curved surface may cause distortions, possibly affecting diagnoses. This would be true especially if physical linear dimensions of the anatomy are important. Reflections from ambient lights behind the observer also could be a problem with curved displays. The goal of this study was to compare physical and psychophysical performance of a flat-surface display monitor with a traditional curved-surface monitor. Two display monitors with different types of front glass-panel surfaces were evaluated. The first monitor had a traditional curved surface, and the other had a flat surface. Physical measurements included dynamic range, display function, veiling glare, and spatial uniformity. An observer performance study used low-contrast, square-wave patterns to determine just-noticeable differences. Ambient lights were turned off in one condition and on in the other. Physical measurements showed that the display functions were nearly identical, but uniformity, veiling glare, and signal-to-noise-ratio were better for the curved monitor. Observer performance was better overall with the curved monitor, but the degradation in performance between lights off and lights on was greater for the curved than flat monitor. The greater degradation with the lights on could be attributed to more reflections off the curved than the flat monitor. A flat-surface display monitor may be useful for viewing clinical radiographs.

Usefulness of segmented trueFISP cardiac pulse sequence in evaluation of congenital and acquired adult cardiac abnormalities.

OBJECTIVE: The purpose of this study is to compare ultrashort TR, segmented trueFISP (fast imaging with steady-state precession) cine MR imaging with segmented FLASH (fast low-angle shot) cine MR imaging for the detection and characterization of congenital and acquired adult cardiac abnormalities. SUBJECTS AND METHODS: Twenty-five patients with known or clinically suspected cardiac abnormalities were imaged on a 1.5-T scanner. Valve plane movies were obtained in patients with suspected valve morphology or function abnormalities or whose horizontal long-axis images showed jets. For each patient, three radiologists independently compared corresponding matched cine FLASH and trueFISP movies for image quality in evaluating anatomy and function of the great vessels and heart. Image quality was rated on a five-point scale, and data were analyzed using both a Wilcoxon's signed rank test and a repeated-measures analysis of variance. RESULTS: Image quality ratings of trueFISP and FLASH showed a statistically significant difference (F = 58.67; df = 1, 72; p < 0.0001), with the average rating for the trueFISP images being significantly higher (mean rating, 4.1 +/- 0.92) than that for the FLASH images (mean, 3.0 +/- 1.0). However, valve architecture in the aortic valves appeared to be better visualized and was more easily measured in valve plane images with FLASH. No statistically significant differences among the ratings of the interpreters (F = 0.018; df = 2, 72; p = 0.9821) were evident, and, therefore, no suggestion of bias was indicated (F = 0.775; df = 1, 2; p = 0.4645). TrueFISP yielded the correct diagnosis prospectively in 13 (100%) of 13 patients, whereas FLASH yielded the correct diagnosis in 12 (92%) of 13 patients. CONCLUSION: TrueFISP images depict morphologic and functional abnormalities with greater clarity and provide greater diagnostic confidence than FLASH images-and in a fraction of the time. A specific exception is in the assessment of valve leaflet architecture and cross-sectional area calculation (i.e., bicuspid aortic valves); in these evaluations, FLASH maintains a complementary diagnostic imaging role.

Correlation of cerebrovascular reserve as measured by acetazolamide-challenged SPECT with angiographic flow patterns and intra- or extracranial arterial stenosis.

BACKGROUND AND PURPOSE: The ability to identify patients at increased risk for stroke from cerebral hemodynamic ischemia may help guide treatment planning. We tested the correlation between regional cerebrovascular reserve (rCVR) on acetazolamide-challenged single-photon emission CT (SPECT) brain scans and intracranial collateral pathways as well as extra- or intracranial (EC-IC) arterial stenosis on cerebral angiography. METHODS: A retrospective analysis of 27 patients who underwent cerebral angiography and acetazolamide-challenged SPECT brain imaging was performed. With cerebral angiography, the anterior, middle, and posterior cerebral artery (ACA, MCA, PCA) territories were evaluated for patterns of flow, including the ipsilateral carotid or basilar arteries, the circle of Willis collaterals, the EC-IC collaterals, and the leptomeningeal collaterals. With acetazolamide-challenged SPECT, the ACA, MCA, and PCA territories were classified as either showing or not showing evidence of decreased rCVR. Statistical significance was determined by the chi(2) test. RESULTS: Patients with decreased rCVR had significantly greater dependence on either the EC-IC or leptomeningeal collaterals (42%) than did patients without decreased rCVR (7%). Similarly, the cerebral hemispheres with decreased rCVR showed a higher prevalence of 70% or greater stenosis or occlusion of the ipsilateral EC-IC arteries in the anterior circulation (74%) than did hemispheres with no evidence of decreased rCVR (16%), and this difference was also statistically significant. CONCLUSION: Acetazolamide-challenged SPECT brain scanning provides additional information regarding rCVR that is not reliably provided by cerebral angiography.

Fluctuations in service loads in an established telemedicine program.

The goal of this investigation was to determine if there were identifiable patterns in the volume and types of teleconsults provided by an established telemedicine program over an extended period of time. Data from over 3 years of providing telemedicine consults within a university-based telemedicine programs were analyzed to identify trends and points of significant change in service provision. Teleconsult volume over a 40-month period was best fit by a logarithmic transformation of the regression curve that is characteristic of slow but steady growth. Consults have been provided in 53 subspecialties, with an average of 12 different subspecialties each month. Number of subspecialties per month was best fit by a sixth-order polynomial. Teleconsult volume has varied on a monthly basis, but overall volume has increased over time. This program has maintained its initial goal of being a multispecialty provider. Analyzing telemedicine consult data over extended periods of time is especially useful for long-term program evaluation and development of a successful business plan.

Gaze dwell times on acute trauma injuries missed because of satisfaction of search.

RATIONALE AND OBJECTIVES: The authors performed this study to determine whether satisfaction of search (SOS) errors in patients with multiple traumas are caused by faulty visual scanning, faulty recognition, or faulty decision making. MATERIALS AND METHODS: A series of radiographs were obtained in patients with multiple traumas. Radiologists interpreted each series under two experimental conditions: when the first radiograph in the series included a fracture, and when it did not. In the first experiment, the initial radiographs showed nondisplaced fractures of the extremities (minor fractures); in the second experiment, the initial radiographs showed abnormalities of greater clinical importance (major fractures). Each series also included a radiograph with a subtle (test) fracture and a normal radiograph on which detection accuracy was measured. In each experiment, gaze dwell time was recorded as 10 radiologists reviewed images from 10 simulated cases of multiple trauma. RESULTS: An SOS effect could be demonstrated only in the second experiment. Analysis of dwell times showed that search on subsequent radiographs was shortened when the initial radiograph contained a fracture; however, the errors were not based on faulty scanning. CONCLUSION: The SOS effect in musculoskeletal trauma is not caused by faulty scanning. Demonstration of an SOS effect on test fractures with major but not minor additional fractures suggests that detection of other fractures is inversely related to the severity of the detected fracture.

The solitary pulmonary nodule on chest radiography: can we really tell if the nodule is calcified?

OBJECTIVE: This study was designed to assess the ability of radiologists to accurately detect calcification within a solitary pulmonary nodule with chest radiography. MATERIALS AND METHODS: Thirty-five solitary pulmonary nodules that were examined by both posteroanterior and lateral chest radiography and on thin-section CT were retrospectively identified. Fourteen radiologists blinded to the results of CT assessed the nodules for the presence or absence of calcification using chest radiographs alone. The radiologists then assigned one of six values on the basis of their confidence in that assessment. The accuracy and confidence values for each nodule were analyzed on the basis of the presence or absence of calcification as seen on CT. Receiver operating characteristic (ROC) curves were generated. RESULTS: The positive predictive value of a "definitely calcified" assessment was 0.93. Combining all levels of radiologists' confidence, the sensitivity of the chest radiograph in the detection of calcium was 0.50 and the specificity was 0.87. There was no difference in the confidence levels reported between the calcified and noncalcified nodules, and there was no correlation of nodule size with accuracy or confidence level. CONCLUSION: The ability of radiologists to detect calcium in a solitary pulmonary nodule by chest radiography was low, as defined by the ROC data. Of the "definitely calcified" nodules, up to 7% may not be calcified and may be potentially malignant. Without documentation of long-term stability, a low threshold for recommending CT may be appropriate.

Telepathology overview: from concept to implementation.

Telepathology is the practice of pathology at a distance by using video imaging and telecommunications. Significant progress has been made in telepathology. To date, 12 classes of telepathology systems have been engineered. Rapid and ultrarapid virtual slide processors may further expand the range of telepathology applications. Next-generation digital imaging light microscopes, such as miniaturized microscope arrays (MMA), may make virtual slide processing a routine laboratory tool. Diagnostic accuracy of telepathology is comparable with that of conventional light microscopy for most diagnoses. Current telepathology applications include intraoperative frozen sections services, routine surgical pathology services, second opinions, and subspecialty consultations. Three telepathology practice models are discussed: the subspecialty practice (SSP) model; the case triage practice (CTP) model; and the virtual group practice (VGP) model. Human factors influence performance with telepathology. Experience with 500 telepathology cases from multiple organs significantly reduces the video viewing time per case (P < .01). Many technology innovations can be represented as S-curves. After long incubation periods, technology use and/or efficiency may accelerate. Telepathology appears to be following an S-curve for a technical innovation.

Femoral neck buttressing: a radiographic and histologic analysis.

OBJECTIVE: To examine the incidence, radiographic and histologic findings of medial femoral neck buttressing in a consecutive group of patients undergoing total hip arthroplasty. DESIGN: Biomechanical parameters were evaluated on standard anteroposterior pelvic radiographs of 113 patients prior to hip replacement surgery. Demographic information on all patients was reviewed and histologic evaluation was performed on specimens obtained at the time of surgery. RESULTS: The incidence of medial femoral neck buttressing was found to be 50% in a consecutive series of patients undergoing total hip arthroplasty. The incidence was slightly higher in women (56% vs. 41%). Patients with buttressing had increased neck-shaft angles and smaller femoral neck diameters than were seen in patients without buttressing. Histologic evaluation demonstrated that the buttress resulted from deposition bone by the periosteum on the femoral neck in the absence of any evidence of femoral neck fracture. CONCLUSION: It would appear that femoral neck buttressing occurs in response to increased joint reactive forces seen at the hip being transmitted through the femoral neck. The increased joint reactive force can be related to the increased neck shaft angle seen in patients with buttressing.

Evaluation of a digital camera for acquiring radiographic images for telemedicine applications.

Many rural sites cannot afford a digitizer to digitize radiographic films and transmit them via a telemedicine network for review by a radiology specialist. This project tested the feasibility of using a consumer digital still camera to photograph radiographic images and transmit them via a telemedicine network to a consulting hub site. In this study, the feasibility of using a digital camera to photograph plain film radiographs of 40 bone trauma cases from a rural health center in Arizona was tested. The cases were transmitted to the Arizona Telemedicine Program hub site using a private asynchronous transfer mode network based on T1 carriers. Two orthopedic surgeons and two radiologists reviewed the cases on a color monitor and the original film images. The readers also rated image quality. There were no significant differences in diagnostic accuracy between conventional film and telemedicine reading. Diagnostic agreement between film and monitor viewing was quite high, as was agreement in confidence ratings. Image quality was generally rated as excellent to good in both viewing conditions. Cases that did not correlate well were judged to have poor image quality, or diagnoses were based on photographs that had part of the diagnostic region of interest cropped off. It was determined that a digital still camera can be used effectively in many cases to photograph radiographic images for transmission and viewing via a telemedicine network, as long as adequate views, zoomed in regions of interest, and good quality original films are used in the acquisition process.

Role of faulty decision making in the satisfaction of search effect in chest radiography.

RATIONALE AND OBJECTIVES: The authors performed this study to determine whether defective pattern recognition or defective decision making is more to blame for satisfaction of search (SOS) errors in chest radiography. MATERIALS AND METHODS: Fifty-eight chest radiographs-half of which demonstrated diverse, native abnormalities-were read by 20 observers. The radiographs were read twice, once with and once without the addition of a simulated pulmonary nodule. Observers provided a verbal account of their focus of attention, indicating suspicious features and regions considered during their inspection of the radiograph. Observers also provided a separate account of the abnormalities they would include in a radiologic report. RESULTS: When the authors considered only those reports that did not refer to the simulated nodules, they found no reduction in the area under the proper receiver operating characteristic (ROC) curves in cases that contained nodules. A smaller SOS effect, however, was demonstrated with analysis of events in which the native abnormality was missed in one condition but not the other. Verbal protocols suggested that the SOS errors were mainly caused by recognition failure rather than faulty decision making. CONCLUSION: Describing their focus of attention may have prompted observers to inspect the radiographs in a more deliberate, systematic way, thus reducing the SOS effect. More residual SOS errors were caused by defective pattern recognition than by faulty decision making.