Expert Agreement on the Presence and Spatial Localization of Melanocytic Features in Dermoscopy.

Dermoscopy aids in melanoma detection; however, agreement on dermoscopic features, including those of high clinical relevance, remains poor. In this study, we attempted to evaluate agreement among experts on exemplar images not only for the presence of melanocytic-specific features but also for spatial localization. This was a cross-sectional, multicenter, observational study. Dermoscopy images exhibiting at least 1 of 31 melanocytic-specific features were submitted by 25 world experts as exemplars. Using a web-based platform that allows for image markup of specific contrast-defined regions (superpixels), 20 expert readers annotated 248 dermoscopic images in collections of 62 images. Each collection was reviewed by five independent readers. A total of 4,507 feature observations were performed. Good-to-excellent agreement was found for 14 of 31 features (45.2%), with eight achieving excellent agreement (Gwet's AC >0.75) and seven of them being melanoma-specific features. These features were peppering/granularity (0.91), shiny white streaks (0.89), typical pigment network (0.83), blotch irregular (0.82), negative network (0.81), irregular globules (0.78), dotted vessels (0.77), and blue-whitish veil (0.76). By utilizing an exemplar dataset, a good-to-excellent agreement was found for 14 features that have previously been shown useful in discriminating nevi from melanoma. All images are public (www.isic-archive.com) and can be used for education, scientific communication, and machine learning experiments.

A reinforcement learning model for AI-based decision support in skin cancer.

We investigated whether human preferences hold the potential to improve diagnostic artificial intelligence (AI)-based decision support using skin cancer diagnosis as a use case. We utilized nonuniform rewards and penalties based on expert-generated tables, balancing the benefits and harms of various diagnostic errors, which were applied using reinforcement learning. Compared with supervised learning, the reinforcement learning model improved the sensitivity for melanoma from 61.4% to 79.5% (95% confidence interval (CI): 73.5-85.6%) and for basal cell carcinoma from 79.4% to 87.1% (95% CI: 80.3-93.9%). AI overconfidence was also reduced while simultaneously maintaining accuracy. Reinforcement learning increased the rate of correct diagnoses made by dermatologists by 12.0% (95% CI: 8.8-15.1%) and improved the rate of optimal management decisions from 57.4% to 65.3% (95% CI: 61.7-68.9%). We further demonstrated that the reward-adjusted reinforcement learning model and a threshold-based model outperformed naïve supervised learning in various clinical scenarios. Our findings suggest the potential for incorporating human preferences into image-based diagnostic algorithms.

Computer algorithms show potential for improving dermatologists' accuracy to diagnose cutaneous melanoma: Results of the International Skin Imaging Collaboration 2017.

BACKGROUND: Computer vision has promise in image-based cutaneous melanoma diagnosis but clinical utility is uncertain. OBJECTIVE: To determine if computer algorithms from an international melanoma detection challenge can improve dermatologists' accuracy in diagnosing melanoma. METHODS: In this cross-sectional study, we used 150 dermoscopy images (50 melanomas, 50 nevi, 50 seborrheic keratoses) from the test dataset of a melanoma detection challenge, along with algorithm results from 23 teams. Eight dermatologists and 9 dermatology residents classified dermoscopic lesion images in an online reader study and provided their confidence level. RESULTS: The top-ranked computer algorithm had an area under the receiver operating characteristic curve of 0.87, which was higher than that of the dermatologists (0.74) and residents (0.66) (P < .001 for all comparisons). At the dermatologists' overall sensitivity in classification of 76.0%, the algorithm had a superior specificity (85.0% vs. 72.6%, P = .001). Imputation of computer algorithm classifications into dermatologist evaluations with low confidence ratings (26.6% of evaluations) increased dermatologist sensitivity from 76.0% to 80.8% and specificity from 72.6% to 72.8%. LIMITATIONS: Artificial study setting lacking the full spectrum of skin lesions as well as clinical metadata. CONCLUSION: Accumulating evidence suggests that deep neural networks can classify skin images of melanoma and its benign mimickers with high accuracy and potentially improve human performance.

Segmentation of Both Diseased and Healthy Skin From Clinical Photographs in a Primary Care Setting.

This work presents the first segmentation study of both diseased and healthy skin in standard camera photographs from a clinical environment. Challenges arise from varied lighting conditions, skin types, backgrounds, and pathological states. For study, 400 clinical photographs (with skin segmentation masks) representing various pathological states of skin are retrospectively collected from a primary care network. 100 images are used for training and fine-tuning, and 300 are used for evaluation. This distribution between training and test partitions is chosen to reflect the difficulty in amassing large quantities of labeled data in this domain. A deep learning approach is used, and 3 public segmentation datasets of healthy skin are collected to study the potential benefits of pretraining. Two variants of U-Net are evaluated: U-Net and Dense Residual U-Net. We find that Dense Residual U-Nets have a 7.8% improvement in Jaccard, compared to classical U-Net architectures (0.55 vs. 0.51 Jaccard), for direct transfer, where fine-tuning data is not utilized. However, U-Net outperforms Dense Residual U-Net for both direct training (0.83 vs. 0.80) and fine-tuning (0.89 vs. 0.88). The stark performance improvement with fine-tuning compared to direct transfer and direct training emphasizes both the need for adequate representative data of diseased skin, and the utility of other publicly available data sources for this task.

Results of the 2016 International Skin Imaging Collaboration International Symposium on Biomedical Imaging challenge: Comparison of the accuracy of computer algorithms to dermatologists for the diagnosis of melanoma from dermoscopic images.

BACKGROUND: Computer vision may aid in melanoma detection. OBJECTIVE: We sought to compare melanoma diagnostic accuracy of computer algorithms to dermatologists using dermoscopic images. METHODS: We conducted a cross-sectional study using 100 randomly selected dermoscopic images (50 melanomas, 44 nevi, and 6 lentigines) from an international computer vision melanoma challenge dataset (n = 379), along with individual algorithm results from 25 teams. We used 5 methods (nonlearned and machine learning) to combine individual automated predictions into "fusion" algorithms. In a companion study, 8 dermatologists classified the lesions in the 100 images as either benign or malignant. RESULTS: The average sensitivity and specificity of dermatologists in classification was 82% and 59%. At 82% sensitivity, dermatologist specificity was similar to the top challenge algorithm (59% vs. 62%, P = .68) but lower than the best-performing fusion algorithm (59% vs. 76%, P = .02). Receiver operating characteristic area of the top fusion algorithm was greater than the mean receiver operating characteristic area of dermatologists (0.86 vs. 0.71, P = .001). LIMITATIONS: The dataset lacked the full spectrum of skin lesions encountered in clinical practice, particularly banal lesions. Readers and algorithms were not provided clinical data (eg, age or lesion history/symptoms). Results obtained using our study design cannot be extrapolated to clinical practice. CONCLUSION: Deep learning computer vision systems classified melanoma dermoscopy images with accuracy that exceeded some but not all dermatologists.

Echocardiographic Algorithm for Post-Myocardial Infarction LV Thrombus: A Gatekeeper for Thrombus Evaluation by Delayed Enhancement CMR.

OBJECTIVES: The goal of this study was to determine the prevalence of post-myocardial infarction (MI) left ventricular (LV) thrombus in the current era and to develop an effective algorithm (predicated on echocardiography [echo]) to discern patients warranting further testing for thrombus via delayed enhancement (DE) cardiac magnetic resonance (CMR). BACKGROUND: LV thrombus affects post-MI management. DE-CMR provides thrombus tissue characterization and is a well-validated but an impractical screening modality for all patients after an MI. METHODS: A same-day echo and CMR were performed according to a tailored protocol, which entailed uniform echo contrast (irrespective of image quality) and dedicated DE-CMR for thrombus tissue characterization. RESULTS: A total of 201 patients were studied; 8% had thrombus according to DE-CMR. All thrombi were apically located; 94% of thrombi occurred in the context of a left anterior descending (LAD) infarct-related artery. Although patients with thrombus had more prolonged chest pain and larger MI (p ≤ 0.01), only 18% had aneurysm on echo (cine-CMR 24%). Noncontrast (35%) and contrast (64%) echo yielded limited sensitivity for thrombus on DE-CMR. Thrombus was associated with stepwise increments in basal → apical contractile dysfunction on echo and quantitative cine-CMR; the echo-measured apical wall motion score was higher among patients with thrombus (p < 0.001) and paralleled cine-CMR decrements in apical ejection fraction and peak ejection rates (both p < 0.005). Thrombus-associated decrements in apical contractile dysfunction were significant even among patients with LAD infarction (p < 0.05). The echo-based apical wall motion score improved overall performance (area under the curve 0.89 ± 0.44) for thrombus compared with ejection fraction (area under the curve 0.80 ± 0.61; p = 0.01). Apical wall motion partitions would have enabled all patients with LV thrombus to be appropriately referred for DE-CMR testing (100% sensitivity and negative predictive value), while avoiding further testing in more than one-half (56% to 63%) of patients. CONCLUSIONS: LV thrombus remains common, especially after LAD MI, and can occur even in the absence of aneurysm. Although DE-CMR yielded improved overall thrombus detection, apical wall motion on a noncontrast echocardiogram can be an effective stratification tool to identify patients in whom DE-CMR thrombus assessment is most warranted. (Diagnostic Utility of Contrast Echocardiography for Detection of LV Thrombi Post ST Elevation Myocardial Infarction; NCT00539045).

Left ventricle: fully automated segmentation based on spatiotemporal continuity and myocardium information in cine cardiac magnetic resonance imaging (LV-FAST).

CMR quantification of LV chamber volumes typically and manually defines the basal-most LV, which adds processing time and user-dependence. This study developed an LV segmentation method that is fully automated based on the spatiotemporal continuity of the LV (LV-FAST). An iteratively decreasing threshold region growing approach was used first from the midventricle to the apex, until the LV area and shape discontinued, and then from midventricle to the base, until less than 50% of the myocardium circumference was observable. Region growth was constrained by LV spatiotemporal continuity to improve robustness of apical and basal segmentations. The LV-FAST method was compared with manual tracing on cardiac cine MRI data of 45 consecutive patients. Of the 45 patients, LV-FAST and manual selection identified the same apical slices at both ED and ES and the same basal slices at both ED and ES in 38, 38, 38, and 41 cases, respectively, and their measurements agreed within -1.6 ± 8.7 mL, -1.4 ± 7.8 mL, and 1.0 ± 5.8% for EDV, ESV, and EF, respectively. LV-FAST allowed LV volume-time course quantitatively measured within 3 seconds on a standard desktop computer, which is fast and accurate for processing the cine volumetric cardiac MRI data, and enables LV filling course quantification over the cardiac cycle.

Improved left ventricular mass quantification with partial voxel interpolation: in vivo and necropsy validation of a novel cardiac MRI segmentation algorithm.

BACKGROUND: Cardiac magnetic resonance (CMR) typically quantifies LV mass (LVM) by means of manual planimetry (MP), but this approach is time-consuming and does not account for partial voxel components--myocardium admixed with blood in a single voxel. Automated segmentation (AS) can account for partial voxels, but this has not been used for LVM quantification. This study used automated CMR segmentation to test the influence of partial voxels on quantification of LVM. METHODS AND RESULTS: LVM was quantified by AS and MP in 126 consecutive patients and 10 laboratory animals undergoing CMR. AS yielded both partial voxel (AS(PV)) and full voxel (AS(FV)) measurements. Methods were independently compared with LVM quantified on echocardiography (echo) and an ex vivo standard of LVM at necropsy. AS quantified LVM in all patients, yielding a 12-fold decrease in processing time versus MP (0:21±0:04 versus 4:18±1:02 minutes; P<0.001). AS(FV) mass (136±35 g) was slightly lower than MP (139±35; Δ=3±9 g, P<0.001). Both methods yielded similar proportions of patients with LV remodeling (P=0.73) and hypertrophy (P=1.00). Regarding partial voxel segmentation, AS(PV) yielded higher LVM (159±38 g) than MP (Δ=20±10 g) and AS(FV) (Δ=23±6 g, both P<0.001), corresponding to relative increases of 14% and 17%. In multivariable analysis, magnitude of difference between AS(PV) and AS(FV) correlated with larger voxel size (partial r=0.37, P<0.001) even after controlling for LV chamber volume (r=0.28, P=0.002) and total LVM (r=0.19, P=0.03). Among patients, AS(PV) yielded better agreement with echo (Δ=20±25 g) than did AS(FV) (Δ=43±24 g) or MP (Δ=40±22 g, both P<0.001). Among laboratory animals, AS(PV) and ex vivo results were similar (Δ=1±3 g, P=0.3), whereas AS(FV) (6±3 g, P<0.001) and MP (4±5 g, P=0.02) yielded small but significant differences with LVM at necropsy. CONCLUSIONS: Automated segmentation of myocardial partial voxels yields a 14-17% increase in LVM versus full voxel segmentation, with increased differences correlated with lower spatial resolution. Partial voxel segmentation yields improved CMR agreement with echo and necropsy-verified LVM.

A radial self-calibrated (RASCAL) generalized autocalibrating partially parallel acquisition (GRAPPA) method using weight interpolation.

A generalized autocalibrating partially parallel acquisition (GRAPPA) method for radial k-space sampling is presented that calculates GRAPPA weights without synthesized or acquired calibration data. Instead, GRAPPA weights are fitted to the undersampled data as if they were the calibration data. Because the relative k-space shifts associated with these GRAPPA weights vary for a radial trajectory, new GRAPPA weights can be resampled for arbitrary shifts through interpolation, which are then used to generate missing projections between the acquired projections. The method is demonstrated in phantoms and in abdominal and brain imaging. Image quality is similar to radial GRAPPA using fully sampled calibration data, and improved relative to a previously described self-calibrated radial GRAPPA technique.

Impact of diastolic dysfunction severity on global left ventricular volumetric filling - assessment by automated segmentation of routine cine cardiovascular magnetic resonance.

OBJECTIVES: To examine relationships between severity of echocardiography (echo) -evidenced diastolic dysfunction (DD) and volumetric filling by automated processing of routine cine cardiovascular magnetic resonance (CMR). BACKGROUND: Cine-CMR provides high-resolution assessment of left ventricular (LV) chamber volumes. Automated segmentation (LV-METRIC) yields LV filling curves by segmenting all short-axis images across all temporal phases. This study used cine-CMR to assess filling changes that occur with progressive DD. METHODS: 115 post-MI patients underwent CMR and echo within 1 day. LV-METRIC yielded multiple diastolic indices - E:A ratio, peak filling rate (PFR), time to peak filling rate (TPFR), and diastolic volume recovery (DVR80 - proportion of diastole required to recover 80% stroke volume). Echo was the reference for DD. RESULTS: LV-METRIC successfully generated LV filling curves in all patients. CMR indices were reproducible (< or = 1% inter-reader differences) and required minimal processing time (175 +/- 34 images/exam, 2:09 +/- 0:51 minutes). CMR E:A ratio decreased with grade 1 and increased with grades 2-3 DD. Diastolic filling intervals, measured by DVR80 or TPFR, prolonged with grade 1 and shortened with grade 3 DD, paralleling echo deceleration time (p < 0.001). PFR by CMR increased with DD grade, similar to E/e' (p < 0.001). Prolonged DVR80 identified 71% of patients with echo-evidenced grade 1 but no patients with grade 3 DD, and stroke-volume adjusted PFR identified 67% with grade 3 but none with grade 1 DD (matched specificity = 83%). The combination of DVR80 and PFR identified 53% of patients with grade 2 DD. Prolonged DVR80 was associated with grade 1 (OR 2.79, CI 1.65-4.05, p = 0.001) with a similar trend for grade 2 (OR 1.35, CI 0.98-1.74, p = 0.06), whereas high PFR was associated with grade 3 (OR 1.14, CI 1.02-1.25, p = 0.02) DD. CONCLUSIONS: Automated cine-CMR segmentation can discern LV filling changes that occur with increasing severity of echo-evidenced DD. Impaired relaxation is associated with prolonged filling intervals whereas restrictive filling is characterized by increased filling rates.

Respiratory and cardiac self-gated free-breathing cardiac CINE imaging with multiecho 3D hybrid radial SSFP acquisition.

A respiratory and cardiac self-gated free-breathing three-dimensional cine steady-state free precession imaging method using multiecho hybrid radial sampling is presented. Cartesian mapping of the k-space center along the slice encoding direction provides intensity-weighted position information, from which both respiratory and cardiac motions are derived. With in plan radial sampling acquired at every pulse repetition time, no extra scan time is required for sampling the k-space center. Temporal filtering based on density compensation is used for radial reconstruction to achieve high signal-to-noise ratio and contrast-to-noise ratio. High correlation between the self-gating signals and external gating signals is demonstrated. This respiratory and cardiac self-gated, free-breathing, three-dimensional, radial cardiac cine imaging technique provides image quality comparable to that acquired with the multiple breath-hold two-dimensional Cartesian steady-state free precession technique in short-axis, four-chamber, and two-chamber orientations. Functional measurements from the three-dimensional cardiac short axis cine images are found to be comparable to those obtained using the standard two-dimensional technique.

Rapid and accurate left ventricular chamber quantification using a novel CMR segmentation algorithm: a clinical validation study.

PURPOSE: To evaluate the clinical performance of a novel automated left ventricle (LV) segmentation algorithm (LV-METRIC) that involves no geometric assumptions. MATERIALS AND METHODS: LV-METRIC and manual tracing (MT) were used independently to quantify LV volumes and LVEF (ejection fraction) for 151 consecutive patients who underwent cine-CMR (steady-state free precession). Phase contrast imaging was used to independently measure stroke volume. RESULTS: LV-METRIC was successful in all cases. Mean LVEF was within 1 point of MT (Delta 0.6 +/- 2.3%, P < 0.05), with smaller differences among patients with (0.5 +/- 2.5%) versus those without (0.9 +/- 2.3%; P = 0.01) advanced systolic dysfunction (LVEF

Automatic left ventricle segmentation using iterative thresholding and an active contour model with adaptation on short-axis cardiac MRI.

An automatic left ventricle (LV) segmentation algorithm is presented for quantification of cardiac output and myocardial mass in clinical practice. The LV endocardium is first segmented using region growth with iterative thresholding by detecting the effusion into the surrounding myocardium and tissues. Then the epicardium is extracted using the active contour model guided by the endocardial border and the myocardial signal information estimated by iterative thresholding. This iterative thresholding and active contour model with adaptation (ITHACA) algorithm was compared to manual tracing used in clinical practice and the commercial MASS Analysis software (General Electric) in 38 patients, with Institutional Review Board (IRB) approval. The ITHACA algorithm provided substantial improvement over the MASS software in defining myocardial borders. The ITHACA algorithm agreed well with manual tracing with a mean difference of blood volume and myocardial mass being 2.9 +/- 6.2 mL (mean +/- standard deviation) and -0.9 +/- 16.5 g, respectively. The difference was smaller than the difference between manual tracing and the MASS software (approximately -20.0 +/- 6.9 mL and -1.0 +/- 20.2 g, respectively). These experimental results support that the proposed ITHACA segmentation is accurate and useful for clinical practice.

Automated segmentation of routine clinical cardiac magnetic resonance imaging for assessment of left ventricular diastolic dysfunction.

BACKGROUND: Cardiac magnetic resonance (CMR) is established for assessment of left ventricular (LV) systolic function but has not been widely used to assess diastolic function. This study tested performance of a novel CMR segmentation algorithm (LV-METRIC) for automated assessment of diastolic function. METHODS AND RESULTS: A total of 101 patients with normal LV systolic function underwent CMR and echocardiography (echo) within 7 days. LV-METRIC generated LV filling profiles via automated segmentation of contiguous short-axis images (204+/-39 images, 2:04+/-0:53 minutes). Diastolic function by CMR was assessed via early:atrial filling ratios, peak diastolic filling rate, time to peak filling rate, and a novel index-diastolic volume recovery (DVR), calculated as percent diastole required for recovery of 80% stroke volume. Using an echo standard, patients with versus without diastolic dysfunction had lower early:atrial filling ratios, longer time to peak filling rate, lower stroke volume-adjusted peak diastolic filling rate, and greater DVR (all P<0.05). Prevalence of abnormal CMR filling indices increased in relation to clinical symptoms classified by New York Heart Association functional class (P=0.04) or dyspnea (P=0.006). Among all parameters tested, DVR yielded optimal performance versus echo (area under the curve: 0.87+/-0.04, P<0.001). Using a 90% specificity cutoff, DVR yielded 74% sensitivity for diastolic dysfunction. In multivariate analysis, DVR (odds ratio, 1.82; 95% CI, 1.13 to 2.57; P=0.02) was independently associated with echo-evidenced diastolic dysfunction after controlling for age, hypertension, and LV mass (chi(2)=73.4, P<0.001). CONCLUSIONS: Automated CMR segmentation can provide LV filling profiles that may offer insight into diastolic dysfunction. Patients with diastolic dysfunction have prolonged diastolic filling intervals, which are associated with echo-evidenced diastolic dysfunction independent of clinical and imaging variables.

Left ventricle: automated segmentation by using myocardial effusion threshold reduction and intravoxel computation at MR imaging.

UNLABELLED: This retrospective analysis of existing patient data had institutional review board approval and was performed in compliance with HIPAA. No informed consent was required. The purpose of the study was to develop and validate an algorithm for automated segmentation of the left ventricular (LV) cavity that accounts for papillary and/or trabecular muscles and partial voxels in cine magnetic resonance (MR) images, an algorithm called LV Myocardial Effusion Threshold Reduction with Intravoxel Computation (LV-METRIC). The algorithm was validated in biologic phantoms, and its results were compared with those of manual tracing, as well as those of a commercial automated segmentation software (MASS [MR Analytical Software System]), in 38 subjects. LV-METRIC accuracy in vitro was 98.7%. Among the 38 subjects studied, LV-METRIC and MASS ejection fraction estimations were highly correlated with manual tracing (R(2) = 0.97 and R(2) = 0.95, respectively). Ventricular volume estimations were smaller with LV-METRIC and larger with MASS than those calculated by using manual tracing, though all results were well correlated (R(2) = 0.99). LV-METRIC volume measurements without partial voxel interpolation were statistically equivalent to manual tracing results (P > .05). LV-METRIC had reduced intraobserver and interobserver variability compared with other methods. MASS required additional manual intervention in 58% of cases, whereas LV-METRIC required no additional corrections. LV-METRIC reliably and reproducibly measured LV volumes. SUPPLEMENTAL MATERIAL: http://radiology.rsnajnls.org/cgi/content/full/248/3/1004/DC1.