Random expert sampling for deep learning segmentation of acute ischemic stroke on non-contrast CT.

BACKGROUND: Outlining acutely infarcted tissue on non-contrast CT is a challenging task for which human inter-reader agreement is limited. We explored two different methods for training a supervised deep learning algorithm: one that used a segmentation defined by majority vote among experts and another that trained randomly on separate individual expert segmentations. METHODS: The data set consisted of 260 non-contrast CT studies in 233 patients with acute ischemic stroke recruited from the multicenter DEFUSE 3 (Endovascular Therapy Following Imaging Evaluation for Ischemic Stroke 3) trial. Additional external validation was performed using 33 patients with matched stroke onset times from the University Hospital Lausanne. A benchmark U-Net was trained on the reference annotations of three experienced neuroradiologists to segment ischemic brain tissue using majority vote and random expert sampling training schemes. The median of volume, overlap, and distance segmentation metrics were determined for agreement in lesion segmentations between (1) three experts, (2) the majority model and each expert, and (3) the random model and each expert. The two sided Wilcoxon signed rank test was used to compare performances (1) to 2) and (1) to (3). We further compared volumes with the 24 hour follow-up diffusion weighted imaging (DWI, final infarct core) and correlations with clinical outcome (modified Rankin Scale (mRS) at 90 days) with the Spearman method. RESULTS: The random model outperformed the inter-expert agreement ((1) to (2)) and the majority model ((1) to (3)) (dice 0.51±0.04 vs 0.36±0.05 (P<0.0001) vs 0.45±0.05 (P<0.0001)). The random model predicted volume correlated with clinical outcome (0.19, P<0.05), whereas the median expert volume and majority model volume did not. There was no significant difference when comparing the volume correlations between random model, median expert volume, and majority model to 24 hour follow-up DWI volume (P>0.05, n=51). CONCLUSION: The random model for ischemic injury delineation on non-contrast CT surpassed the inter-expert agreement ((1) to (2)) and the performance of the majority model ((1) to (3)). We showed that the random model volumetric measures of the model were consistent with 24 hour follow-up DWI.

Evaluation of the quality and the productivity of neuroradiological reading of multiple sclerosis follow-up MRI scans using an intelligent automation software.

PURPOSE: The assessment of multiple sclerosis (MS) lesions on follow-up magnetic resonance imaging (MRI) is tedious, time-consuming, and error-prone. Automation of low-level tasks could enhance the radiologist in this work. We evaluate the intelligent automation software Jazz in a blinded three centers study, for the assessment of new, slowly expanding, and contrast-enhancing MS lesions. METHODS: In three separate centers, 117 MS follow-up MRIs were blindly analyzed on fluid attenuated inversion recovery (FLAIR), pre- and post-gadolinium T1-weighted images using Jazz by 2 neuroradiologists in each center. The reading time was recorded. The ground truth was defined in a second reading by side-by-side comparison of both reports from Jazz and the standard clinical report. The number of described new, slowly expanding, and contrast-enhancing lesions described with Jazz was compared to the lesions described in the standard clinical report. RESULTS: A total of 96 new lesions from 41 patients and 162 slowly expanding lesions (SELs) from 61 patients were described in the ground truth reading. A significantly larger number of new lesions were described using Jazz compared to the standard clinical report (63 versus 24). No SELs were reported in the standard clinical report, while 95 SELs were reported on average using Jazz. A total of 4 new contrast-enhancing lesions were found in all reports. The reading with Jazz was very time efficient, taking on average 2min33s ± 1min0s per case. Overall inter-reader agreement for new lesions between the readers using Jazz was moderate for new lesions (Cohen kappa = 0.5) and slight for SELs (0.08). CONCLUSION: The quality and the productivity of neuroradiological reading of MS follow-up MRI scans can be significantly improved using the dedicated software Jazz.

Artificial intelligence tools in clinical neuroradiology: essential medico-legal aspects.

Commercial software based on artificial intelligence (AI) is entering clinical practice in neuroradiology. Consequently, medico-legal aspects of using Software as a Medical Device (SaMD) become increasingly important. These medico-legal issues warrant an interdisciplinary approach and may affect the way we work in daily practice. In this article, we seek to address three major topics: medical malpractice liability, regulation of AI-based medical devices, and privacy protection in shared medical imaging data, thereby focusing on the legal frameworks of the European Union and the USA. As many of the presented concepts are very complex and, in part, remain yet unsolved, this article is not meant to be comprehensive but rather thought-provoking. The goal is to engage clinical neuroradiologists in the debate and equip them to actively shape these topics in the future.

Deep learning intravoxel incoherent motion modeling: Exploring the impact of training features and learning strategies.

PURPOSE: The development of advanced estimators for intravoxel incoherent motion (IVIM) modeling is often motivated by a desire to produce smoother parameter maps than least squares (LSQ). Deep neural networks show promise to this end, yet performance may be conditional on a myriad of choices regarding the learning strategy. In this work, we have explored potential impacts of key training features in unsupervised and supervised learning for IVIM model fitting. METHODS: Two synthetic data sets and one in-vivo data set from glioma patients were used in training of unsupervised and supervised networks for assessing generalizability. Network stability for different learning rates and network sizes was assessed in terms of loss convergence. Accuracy, precision, and bias were assessed by comparing estimations against ground truth after using different training data (synthetic and in vivo). RESULTS: A high learning rate, small network size, and early stopping resulted in sub-optimal solutions and correlations in fitted IVIM parameters. Extending training beyond early stopping resolved these correlations and reduced parameter error. However, extensive training resulted in increased noise sensitivity, where unsupervised estimates displayed variability similar to LSQ. In contrast, supervised estimates demonstrated improved precision but were strongly biased toward the mean of the training distribution, resulting in relatively smooth, yet possibly deceptive parameter maps. Extensive training also reduced the impact of individual hyperparameters. CONCLUSION: Voxel-wise deep learning for IVIM fitting demands sufficiently extensive training to minimize parameter correlation and bias for unsupervised learning, or demands a close correspondence between the training and test sets for supervised learning.

Correlating volumetric and linear measurements of brain metastases on MRI scans using intelligent automation software: a preliminary study.

PURPOSE: The Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) working group proposed a guide for treatment responses for BMs by utilizing the longest diameter; however, despite recognizing that many patients with BMs have sub-centimeter lesions, the group referred to these lesions as unmeasurable due to issues with repeatability and interpretation. In light of RANO-BM recommendations, we aimed to correlate linear and volumetric measurements in sub-centimeter BMs on contrast-enhanced MRI using intelligent automation software. METHODS: In this retrospective study, patients with BMs scanned with MRI between January 1, 2018, and December 31, 2021, were screened. Inclusion criteria were: (1) at least one sub-centimeter BM with an integer millimeter-longest diameter was noted in the MRI report; (2) patients were a minimum of 18 years of age; (3) patients with available pre-treatment three-dimensional T1-weighted spoiled gradient-echo MRI scan. The screening was terminated when there were 20 lesions in each group. Lesion volumes were measured with the help of intelligent automation software Jazz (AI Medical, Zollikon, Switzerland) by two readers. The Kruskal-Wallis test was used to compare volumetric differences. RESULTS: Our study included 180 patients. The agreement for volumetric measurements was excellent between the two readers. The volumes of the following groups were not significantly different: 1-2 mm, 1-3 mm, 1-4 mm, 2-3 mm, 2-4 mm, 3-4 mm, 3-5 mm, 4-5 mm, 5-6 mm, 5-7 mm, 6-7 mm, 6-8 mm, 6-9 mm, 7-8 mm, 7-9 mm, 8-9 mm. CONCLUSION: Our findings indicate that the largest diameter of a lesion may not accurately represent its volume. Additional research is required to determine which method is superior for measuring radiologic response to therapy and which parameter correlates best with clinical improvement or deterioration.

Deep Learning for Detecting Brain Metastases on MRI: A Systematic Review and Meta-Analysis.

Since manual detection of brain metastases (BMs) is time consuming, studies have been conducted to automate this process using deep learning. The purpose of this study was to conduct a systematic review and meta-analysis of the performance of deep learning models that use magnetic resonance imaging (MRI) to detect BMs in cancer patients. A systematic search of MEDLINE, EMBASE, and Web of Science was conducted until 30 September 2022. Inclusion criteria were: patients with BMs; deep learning using MRI images was applied to detect the BMs; sufficient data were present in terms of detective performance; original research articles. Exclusion criteria were: reviews, letters, guidelines, editorials, or errata; case reports or series with less than 20 patients; studies with overlapping cohorts; insufficient data in terms of detective performance; machine learning was used to detect BMs; articles not written in English. Quality Assessment of Diagnostic Accuracy Studies-2 and Checklist for Artificial Intelligence in Medical Imaging was used to assess the quality. Finally, 24 eligible studies were identified for the quantitative analysis. The pooled proportion of patient-wise and lesion-wise detectability was 89%. Articles should adhere to the checklists more strictly. Deep learning algorithms effectively detect BMs. Pooled analysis of false positive rates could not be estimated due to reporting differences.

The R-AI-DIOLOGY checklist: a practical checklist for evaluation of artificial intelligence tools in clinical neuroradiology.

Artificial intelligence (AI)-based tools are gradually blending into the clinical neuroradiology practice. Due to increasing complexity and diversity of such AI tools, it is not always obvious for the clinical neuroradiologist to capture the technical specifications of these applications, notably as commercial tools very rarely provide full details. The clinical neuroradiologist is thus confronted with the increasing dilemma to base clinical decisions on the output of AI tools without knowing in detail what is happening inside the "black box" of those AI applications. This dilemma is aggravated by the fact that currently, no established and generally accepted rules exist concerning best clinical practice and scientific and clinical validation nor for the medico-legal consequences in cases of wrong diagnoses. The current review article provides a practical checklist of essential points, intended to aid the user to identify and double-check necessary aspects, although we are aware that not all this information may be readily available at this stage, even for certified and commercially available AI tools. Furthermore, we therefore suggest that the developers of AI applications provide this information.

Signal to noise and b-value analysis for optimal intra-voxel incoherent motion imaging in the brain.

Intravoxel incoherent motion (IVIM) is a method that can provide quantitative information about perfusion in the human body, in vivo, and without contrast agent. Unfortunately, the IVIM perfusion parameter maps are known to be relatively noisy in the brain, in particular for the pseudo-diffusion coefficient, which might hinder its potential broader use in clinical applications. Therefore, we studied the conditions to produce optimal IVIM perfusion images in the brain. IVIM imaging was performed on a 3-Tesla clinical system in four healthy volunteers, with 16 b values 0, 10, 20, 40, 80, 110, 140, 170, 200, 300, 400, 500, 600, 700, 800, 900 s/mm2, repeated 20 times. We analyzed the noise characteristics of the trace images as a function of b-value, and the homogeneity of the IVIM parameter maps across number of averages and sub-sets of the acquired b values. We found two peaks of noise of the trace images as function of b value, one due to thermal noise at high b-value, and one due to physiological noise at low b-value. The selection of b value distribution was found to have higher impact on the homogeneity of the IVIM parameter maps than the number of averages. Based on evaluations, we suggest an optimal b value acquisition scheme for a 12 min scan as 0 (7), 20 (4), 140 (19), 300 (9), 500 (19), 700 (1), 800 (4), 900 (1) s/mm2.

Multi-Modal Segmentation of 3D Brain Scans Using Neural Networks.

Anatomical segmentation of brain scans is highly relevant for diagnostics and neuroradiology research. Conventionally, segmentation is performed on T 1-weighted MRI scans, due to the strong soft-tissue contrast. In this work, we report on a comparative study of automated, learning-based brain segmentation on various other contrasts of MRI and also computed tomography (CT) scans and investigate the anatomical soft-tissue information contained in these imaging modalities. A large database of in total 853 MRI/CT brain scans enables us to train convolutional neural networks (CNNs) for segmentation. We benchmark the CNN performance on four different imaging modalities and 27 anatomical substructures. For each modality we train a separate CNN based on a common architecture. We find average Dice scores of 86.7 ± 4.1% (T 1-weighted MRI), 81.9 ± 6.7% (fluid-attenuated inversion recovery MRI), 80.8 ± 6.6% (diffusion-weighted MRI) and 80.7 ± 8.2% (CT), respectively. The performance is assessed relative to labels obtained using the widely-adopted FreeSurfer software package. The segmentation pipeline uses dropout sampling to identify corrupted input scans or low-quality segmentations. Full segmentation of 3D volumes with more than 2 million voxels requires <1s of processing time on a graphical processing unit.

Optimizing Deep Learning Algorithms for Segmentation of Acute Infarcts on Non-Contrast Material-enhanced CT Scans of the Brain Using Simulated Lesions.

PURPOSE: To test the efficacy of lesion segmentation using a deep learning algorithm on non-contrast material-enhanced CT (NCCT) images with synthetic lesions resembling acute infarcts. MATERIALS AND METHODS: In this retrospective study, 40 diffusion-weighted imaging (DWI) lesions in patients with acute stroke (median age, 69 years; range, 62-76 years; 17 women; screened between 2011 and 2017) were coregistered to 40 normal NCCT scans (median age, 70 years; range, 55-76 years; 25 women; screened between 2008 and 2011), which produced 640 combinations of DWI-NCCT with and without lesions for training (n = 420), validation (n = 110), and testing (n = 110). The signal intensity on the NCCT scans was depressed by 4 HU (a 13% drop) in the region of the diffusion-weighted lesion. Two U-Net architectures (standard and symmetry aware) were trained with two different training strategies. One was a naive strategy, in which the model started training with random coefficients. The other was a progressive strategy, which started with coefficients derived from a model trained on a dataset with lesions that were depressed by 10 HU. The Dice scores from the two architectures and training strategies were compared from the test dataset. RESULTS: Dice scores of symmetry-aware U-Nets were 25% higher than those of standard U-Nets (median, 0.49 vs 0.65; P < .001). Use of a progressive training strategy had no clear effect on model performance. CONCLUSION: Symmetry-aware U-Nets offer promise for segmentation of acute stroke lesions on NCCT scans.Keywords: Adults, CT-Quantitative, StrokeSupplemental material is available for this article.© RSNA, 2021.

Bayesian inference using hierarchical and spatial priors for intravoxel incoherent motion MR imaging in the brain: Analysis of cancer and acute stroke.

The intravoxel incoherent motion (IVIM) model allows to map diffusion (D) and perfusion-related parameters (F and D*). Parameter estimation is, however, error-prone due to the non-linearity of the signal model, the limited signal-to-noise ratio (SNR) and the small volume fraction of perfusion in the in-vivo brain. In the present work, the performance of Bayesian inference was examined in the presence of brain pathologies characterized by hypo- and hyperperfusion. In particular, a hierarchical and a spatial prior were combined. Performance was compared relative to conventional segmented least squares regression, hierarchical prior only (non-segmented and segmented data likelihoods) and a deep learning approach. Realistic numerical brain IVIM simulations were conducted to assess errors relative to ground truth. In-vivo, data of 11 central nervous system cancer patients and 9 patients with acute stroke were acquired. The proposed method yielded reduced error in simulations for both the cancer and acute stroke scenarios compared to other methods across the whole investigated SNR range. The contrast-to-noise ratio of the proposed method was better or on par compared to the other techniques in-vivo. The proposed Bayesian approach hence improves IVIM parameter estimation in brain cancer and acute stroke.

Realistic generation of diffusion-weighted magnetic resonance brain images with deep generative models.

We study two state of the art deep generative networks, the Introspective Variational Autoencoder and the Style-Based Generative Adversarial Network, for the generation of new diffusion-weighted magnetic resonance images. We show that high quality, diverse and realistic-looking images, as evaluated by external neuroradiologists blinded to the whole study, can be synthesized using these deep generative models. We evaluate diverse metrics with respect to quality and diversity of the generated synthetic brain images. These findings show that generative models could qualify as a method for data augmentation in the medical field, where access to large image database is in many aspects restricted.

Simulation of intravoxel incoherent perfusion signal using a realistic capillary network of a mouse brain.

PURPOSE: To simulate the intravoxel incoherent perfusion magnetic resonance magnitude signal from the motion of blood particles in three realistic vascular network graphs from a mouse brain. METHODS: In three networks generated from the cortex of a mouse scanned by two-photon laser microscopy, blood flow in each vessel was simulated using Poiseuille's law. The trajectories, flow speeds and phases acquired by a fixed number of simulated blood particles during a Stejskal-Tanner bipolar pulse gradient scheme were computed. The resulting magnitude signal was obtained by integrating all phases and the pseudo-diffusion coefficient D* was estimated by fitting an exponential signal decay. To better understand the anatomical source of the intravoxel incoherent motion (IVIM) perfusion signal, the above was repeated restricting the simulation to various types of vessel. RESULTS: The characteristics of the three microvascular networks were respectively vessel lengths (mean ± std. dev.) 67.2 ± 53.6 µm, 59.8 ± 46.2 µm and 64.5 ± 50.9 µm, diameters 6.0 ± 3.5 µm, 5.7 ± 3.6 µm and 6.1 ± 3.7 µm and simulated blood velocity 0.9 ± 1.7 µm/ms, 1.4 ± 2.5 µm/ms and 0.7 ± 2.1 µm/ms. Exponential fitting of the simulated signal decay as a function of b-value resulted in the following D*-values [10-3 mm2 /s]: 31.7, 40.4 and 33.4. The signal decay for low b-values was the largest in the larger vessels, but the smaller vessels and the capillaries accounted for more of the total volume of the networks. CONCLUSION: This simulation improves the theoretical understanding of the IVIM perfusion estimation method by directly linking the MR IVIM perfusion signal to an ultra-high resolution measurement of the microvascular network and a realistic blood flow simulation.

Measuring Perfusion: Intravoxel Incoherent Motion MR Imaging.

The signal acquired in vivo using a diffusion-weighted MR imaging (DWI) sequence is influenced by blood motion in the tissue. This means that perfusion information from a DWI sequence can be obtained in addition to thermal diffusion, if the appropriate sequence parameters and postprocessing methods are applied. This is commonly regrouped under the denomination intravoxel incoherent motion (IVIM) perfusion MR imaging. Of relevance, the perfusion information acquired with IVIM is essentially local, quantitative and acquired without intravenous injection of contrast media. The aim of this work is to review the IVIM method and its clinical applications.

Effect of Sex on Clinical Outcome and Imaging after Endovascular Treatment of Large-Vessel Ischemic Stroke.

BACKGROUND AND PURPOSE: It is unclear if sex differences explain some of the variability in the outcomes of stroke patients who undergo endovascular treatment (EVT). In this study we assess the effect of sex on radiological and functional outcomes in EVT-treated acute stroke patients and determine if differences in baseline perfusion status between men and women might account for differences in outcomes. METHODS: We included patients from the CRISP (Computed tomographic perfusion to Predict Response to Recanalization in ischemic stroke) study, a prospective cohort study of acute stroke patients who underwent EVT up to 18 hours after last seen well. We designed ordinal regression and univariable and multivariable regression models to examine the association between sex and infarct growth, final infarct volume and 90-day mRS score. RESULTS: We included 198 patients. At baseline, women had smaller perfusion lesions, more often had a target mismatch perfusion profile, and had better collateral perfusion. Women experienced less ischemic core growth (median 15 mL vs. 29 mL, p < 0.01) and had smaller final infarct volumes (median 26 mL vs. 50 mL, p < 0.01). Female sex was associated with a favorable shift on the modified Rankin Scale (adjusted cOR 1.79 [1.04 - 3.08; p = 0.04]) and lower odds of severe disability or death (adjusted OR 0.29 [0.10 - 0.81]; p = 0.02). CONCLUSIONS: The results suggest that women have better collaterals and, therefore, more often exhibit a favorable imaging profile on baseline imaging, experience less lesion growth, and have better clinical outcomes following endovascular therapy.

Evaluation of the clinical utility of maximum intensity projections of 3D contrast-enhanced, T1-weighted imaging for the detection of brain metastases.

BACKGROUND: To visualize and assess brain metastases on magnetic resonance imaging, radiologists face an ever-increasing pressure to perform faster and more efficiently. The usage of maximum intensity projections (MIPs) of contrast-enhanced T1-weighed (T1ce) magnetization-prepared rapid acquisition with gradient echo (MP-RAGE) images proposes to increase reading efficiency by increasing lesion conspicuity while reducing in the number of images to be reviewed. AIM: To assess if MIPs save reading time and achieve the same level of diagnostic accuracy as standard 1 mm T1ce images for the detection of brain metastases. METHODS: Forty-four patients were included in this retrospective study. Axial reformations of T1ce MP-RAGE (TR/TE = 2300/2.25 ms, resolution = 1 mm3 ) images were analyzed and post-processed into 5 and 10 mm MIPs. Two readers evaluated the randomly assorted images and recorded reading time. Reading time differences were analyzed using the Wilcoxon test, and inter-reader statistics were performed using Bland-Altman plots. RESULTS: About 22.5 61.2 s/study and 43.8 ± 159.9 s/study were saved using 5 and 10 mm MIPs, respectively. Combined average sensitivity was 92.0% for 5 mm MIPs and 86.3% for 10 mm MIPs compared to standard 1 mm axial slices, with an average rate of 0.98 and 0.57 false positives per study, respectively CONCLUSION: While 5 mm and 10 mm T1ce MP-RAGE MIPs showed a clinical benefit in reducing reading times for evaluation of brain metastases, they should be used in conjunction with standard 1 mm images for best sensitivity and specificity, a practice which possibly annuls their benefit.

Tilt-Corrected Region Boundaries May Enhance the Alberta Stroke Program Early Computed Tomography Score for Less Experienced Raters.

BACKGROUND: The Alberta Stroke Program Early Computed Tomography Score (ASPECTS) is widely used to quantify early ischemic changes in the anterior circulation but has limited inter-rater reliability. AIMS: We investigated whether application of 3-dimensional boundaries outlining the ASPECTS regions improves inter-rater reliability and accuracy. METHODS: We included all patients from our DEFUSE 2 database who had a pretreatment noncontrast computed tomography scan (NCCT) of acceptable quality. Six raters (2 neuroradiologists, 2 vascular neurologists, and 2 neurology residents) scored ASPECTS of each NCCT without ("CT-native") and with the superimposed boundary template ("CT-template"). Gold-standard ASPECTS were generated by the 2 neuroradiologists through joint adjudication. Inter-rater reliability and accuracy were assessed using the intraclass correlation coefficient (ICC) for full-scale agreements and Gwet's AC1 for dichotomized (ASPECTS 0-6 vs 7-10) agreements. RESULTS: Eighty-two patients were included. Inter-rater reliability improved with higher training level for both CT-native (ICC = .15, .31, .54 for residents, neurologists, and radiologists, respectively) and CT-template (ICC = .18, .33, .56). Use of the boundary template improved correlation with the gold-standard for one resident on full-scale agreement (ICC increased from .01 to .31, P = .01) and another resident on dichotomized agreement (AC1 increased from .36 to .64, P = .01), but resulted in no difference for other raters. The template did not improve ICC between raters of the same training level. CONCLUSIONS: Inter-rater reliability of ASPECTS improves with physician training level. Standardized display of ASPECTS region boundaries on NCCT does not improve inter-rater reliability but may improve accuracy for some less experienced raters.

Improved Segmentation and Detection Sensitivity of Diffusion-weighted Stroke Lesions with Synthetically Enhanced Deep Learning.

PURPOSE: To compare the segmentation and detection performance of a deep learning model trained on a database of human-labeled clinical stroke lesions on diffusion-weighted (DW) images to a model trained on the same database enhanced with synthetic stroke lesions. MATERIALS AND METHODS: In this institutional review board-approved study, a stroke database of 962 cases (mean patient age ± standard deviation, 65 years ± 17; 255 male patients; 449 scans with DW positive stroke lesions) and a normal database of 2027 patients (mean age, 38 years ± 24; 1088 female patients) were used. Brain volumes with synthetic stroke lesions on DW images were produced by warping the relative signal increase of real strokes to normal brain volumes. A generic three-dimensional (3D) U-Net was trained on four different databases to generate four different models: (a) 375 neuroradiologist-labeled clinical DW positive stroke cases (CDB); (b) 2000 synthetic cases (S2DB); (c) CDB plus 2000 synthetic cases (CS2DB); and (d) CDB plus 40 000 synthetic cases (CS40DB). The models were tested on 20% (n = 192) of the cases of the stroke database, which were excluded from the training set. Segmentation accuracy was characterized using Dice score and lesion volume of the stroke segmentation, and statistical significance was tested using a paired two-tailed Student t test. Detection sensitivity and specificity were compared with labeling done by three neuroradiologists. RESULTS: The performance of the 3D U-Net model trained on the CS40DB (mean Dice score, 0.72) was better than models trained on the CS2DB (Dice score, 0.70; P < .001) or the CDB (Dice score, 0.65; P < .001). The deep learning model (CS40DB) was also more sensitive (91% [95% confidence interval {CI}: 89%, 93%]) than each of the three human readers (human reader 3, 84% [95% CI: 81%, 87%]; human reader 1, 78% [95% CI: 75%, 81%]; human reader 2, 79% [95% CI: 76%, 82%]), but was less specific (75% [95% CI: 72%, 78%]) than each of the three human readers (human reader 3, 96% [95% CI: 94%, 98%]; human reader 1, 92% [95% CI: 90%, 94%]; human reader 2, 89% [95% CI: 86%, 91%]). CONCLUSION: Deep learning training for segmentation and detection of stroke lesions on DW images was significantly improved by enhancing the training set with synthetic lesions.Supplemental material is available for this article.© RSNA, 2020.