The Impact of Multi-Institution Datasets on the Generalizability of Machine Learning Prediction Models in the ICU.

OBJECTIVES: To evaluate the transferability of deep learning (DL) models for the early detection of adverse events to previously unseen hospitals. DESIGN: Retrospective observational cohort study utilizing harmonized intensive care data from four public datasets. SETTING: ICUs across Europe and the United States. PATIENTS: Adult patients admitted to the ICU for at least 6 hours who had good data quality. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Using carefully harmonized data from a total of 334,812 ICU stays, we systematically assessed the transferability of DL models for three common adverse events: death, acute kidney injury (AKI), and sepsis. We tested whether using more than one data source and/or algorithmically optimizing for generalizability during training improves model performance at new hospitals. We found that models achieved high area under the receiver operating characteristic (AUROC) for mortality (0.838-0.869), AKI (0.823-0.866), and sepsis (0.749-0.824) at the training hospital. As expected, AUROC dropped when models were applied at other hospitals, sometimes by as much as -0.200. Using more than one dataset for training mitigated the performance drop, with multicenter models performing roughly on par with the best single-center model. Dedicated methods promoting generalizability did not noticeably improve performance in our experiments. CONCLUSIONS: Our results emphasize the importance of diverse training data for DL-based risk prediction. They suggest that as data from more hospitals become available for training, models may become increasingly generalizable. Even so, good performance at a new hospital still depended on the inclusion of compatible hospitals during training.

Temporal Shape Changes of Pedicle screw-rod Constructs After Lumbar Interbody Fusion.

STUDY DESIGN: Retrospective multicenter study. OBJECTIVE: To examine the shape change of screw-rod constructs over time following short-segment lumbar interbody fusion and to clarify its relationship to clinical characteristics. SUMMARY OF BACKGROUND DATA: No study has focused on the shape change of screw-rod constructs after short-segment fusion and its clinical implications. METHODS: One hundred and eight patients who had single-level lumbar interbody fusion with pedicle screws and cages were enrolled. Three-dimensional (3D) images of screw-rod constructs were generated from baseline CT on the day after surgery and follow-up CT, and were superposed on the right and left side, respectively, using the iterative closest point algorithm. The shape change was quantitatively assessed by computing the median distance between the 3D images, which was defined as the shape change value. Among the five time-course categories of follow-up CT (≤1 month, 2-3 months, 4-6 months, 7-12 months, ≥13 months), the shape change values were compared. The relationships between the shape change values and clinical characteristics, such as age, CT-derived vertebral bone mineral density, screw and rod materials, and postoperative interbody fusion status, cage subsidence, and screw loosening, were evaluated. RESULTS: A total of 237 follow-up CTs were included (≤1 month [34 scans], 2-3 months [33 scans], 4-6 months [80 scans], 7-12 months [48 scans], ≥13 months [42 scans]) because many patients underwent multiple follow-up CTs. There were significant differences in shape change values among the time-course categories (P<0.001 in Kruskal-Wallis test). Most shape changes occurred within 6 months postoperatively, with no significant changes observed at 7 months or more. There were no significant relationships between the shape change values and each clinical characteristic. CONCLUSION: The temporal shape changes of screw-rod constructs following short-segment lumbar interbody fusion progressed up to 6 months after surgery but not significantly thereafter.

Preoperatively-determined Red Distribution Width (RDW) predicts prolonged length of stay after single-level spinal fusion in elderly patients.

Introduction: Elderly patients receiving lumbar fusion surgeries present with a higher risk profile, which necessitates a robust predictor of postoperative outcomes. The Red Distribution Width (RDW) is a preoperative routinely determined parameter that reflects the degree of heterogeneity of red blood cells. Thereby, RDW is associated with frailty in hospital-admitted patients. Research question: This study aims to elucidate the potential of RDW as a frailty biomarker predictive of prolonged hospital stays following elective mono-segmental fusion surgery in elderly patients. Material and methods: In this retrospective study, we included all patients with age over 75 years that were treated via lumbar single-level spinal fusion from 2015 to 2022 at our tertiary medical center. Prolonged length of stay (pLOS) was defined as a length ≥ the 3rd quartile of LOS of all included patients. Classical correlation analysis, Receiver-operating characteristic (ROC) and new machine learning algorithms) were used. Results: A total of 208 patients were included in the present study. The median age was 77 (IQR 75-80) years. The median LOS of the patients was 6 (IQR 5-8) days. The data shows a significant positive correlation between RDW and LOS. RDW is significantly enhanced in the pLOS group. New machine learning approaches with the imputation of multiple variables can enhance the performance to an AUC of 71%. Discussion and conclusion: RDW may serve as a predictor for a pLOS in elderly. These results are compelling because the determination of this frailty biomarker is routinely performed at hospital admission. An improved prognostication of LOS could enable healthcare systems to distribute constrained hospital resources efficiently, fostering evidence-based decision-making processes.

A deep learning analysis of stroke onset time prediction and comparison to DWI-FLAIR mismatch.

INTRODUCTION: When time since stroke onset is unknown, DWI-FLAIR mismatch rating is an established technique for patient stratification. A visible DWI lesion without corresponding parenchymal hyperintensity on FLAIR suggests time since onset of under 4.5 h and thus a potential benefit from intravenous thrombolysis. To improve accuracy and availability of the mismatch concept, deep learning might be able to augment human rating and support decision-making in these cases. METHODS: We used unprocessed DWI and coregistered FLAIR imaging data to train a deep learning model to predict dichotomized time since ischemic stroke onset. We analyzed the performance of Group Convolutional Neural Networks compared to other deep learning methods. Unlabeled imaging data was used for pre-training. Prediction performance of the best deep learning model was compared to the performance of four independent junior and senior raters. Additionally, in cases deemed indeterminable by human raters, model ratings were used to augment human performance. Post-hoc gradient-based explanations were analyzed to gain insights into model predictions. RESULTS: Our best predictive model performed comparably to human raters. Using model ratings in cases deemed indeterminable by human raters improved rating accuracy and interrater agreement for junior and senior ratings. Post-hoc explainability analyses showed that the model localized stroke lesions to derive predictions. DISCUSSION: Our analysis shows that deep learning based clinical decision support has the potential to improve the accessibility of the DWI-FLAIR mismatch concept by supporting patient stratification.

Personalised simulation of hemodynamics in cerebrovascular disease: lessons learned from a study of diagnostic accuracy.

Intracranial atherosclerotic disease (ICAD) poses a significant risk of subsequent stroke but current prevention strategies are limited. Mechanistic simulations of brain hemodynamics offer an alternative precision medicine approach by utilising individual patient characteristics. For clinical use, however, current simulation frameworks have insufficient validation. In this study, we performed the first quantitative validation of a simulation-based precision medicine framework to assess cerebral hemodynamics in patients with ICAD against clinical standard perfusion imaging. In a retrospective analysis, we used a 0-dimensional simulation model to detect brain areas that are hemodynamically vulnerable to subsequent stroke. The main outcome measures were sensitivity, specificity, and area under the receiver operating characteristics curve (ROC AUC) of the simulation to identify brain areas vulnerable to subsequent stroke as defined by quantitative measurements of relative mean transit time (relMTT) from dynamic susceptibility contrast MRI (DSC-MRI). In 68 subjects with unilateral stenosis >70% of the internal carotid artery (ICA) or middle cerebral artery (MCA), the sensitivity and specificity of the simulation were 0.65 and 0.67, respectively. The ROC AUC was 0.68. The low-to-moderate accuracy of the simulation may be attributed to assumptions of Newtonian blood flow, rigid vessel walls, and the use of time-of-flight MRI for geometric representation of subject vasculature. Future simulation approaches should focus on integrating additional patient data, increasing accessibility of precision medicine tools to clinicians, addressing disease burden disparities amongst different populations, and quantifying patient benefit. Our results underscore the need for further improvement of mechanistic simulations of brain hemodynamics to foster the translation of the technology to clinical practice.

Enhancing the prediction for shunt-dependent hydrocephalus after aneurysmal subarachnoid hemorrhage using a machine learning approach.

Early and reliable prediction of shunt-dependent hydrocephalus (SDHC) after aneurysmal subarachnoid hemorrhage (aSAH) may decrease the duration of in-hospital stay and reduce the risk of catheter-associated meningitis. Machine learning (ML) may improve predictions of SDHC in comparison to traditional non-ML methods. ML models were trained for CHESS and SDASH and two combined individual feature sets with clinical, radiographic, and laboratory variables. Seven different algorithms were used including three types of generalized linear models (GLM) as well as a tree boosting (CatBoost) algorithm, a Naive Bayes (NB) classifier, and a multilayer perceptron (MLP) artificial neural net. The discrimination of the area under the curve (AUC) was classified (0.7 ≤ AUC < 0.8, acceptable; 0.8 ≤ AUC < 0.9, excellent; AUC ≥ 0.9, outstanding). Of the 292 patients included with aSAH, 28.8% (n = 84) developed SDHC. Non-ML-based prediction of SDHC produced an acceptable performance with AUC values of 0.77 (CHESS) and 0.78 (SDASH). Using combined feature sets with more complex variables included than those incorporated in the scores, the ML models NB and MLP reached excellent performances, with an AUC of 0.80, respectively. After adding the amount of CSF drained within the first 14 days as a late feature to ML-based prediction, excellent performances were reached in the MLP (AUC 0.81), NB (AUC 0.80), and tree boosting model (AUC 0.81). ML models may enable clinicians to reliably predict the risk of SDHC after aSAH based exclusively on admission data. Future ML models may help optimize the management of SDHC in aSAH by avoiding delays in clinical decision-making.

Artificial Intelligence for Clinical Decision Support in Acute Ischemic Stroke: A Systematic Review.

BACKGROUND: Established randomized trial-based parameters for acute ischemic stroke group patients into generic treatment groups, leading to attempts using various artificial intelligence (AI) methods to directly correlate patient characteristics to outcomes and thereby provide decision support to stroke clinicians. We review AI-based clinical decision support systems in the development stage, specifically regarding methodological robustness and constraints for clinical implementation. METHODS: Our systematic review included full-text English language publications proposing a clinical decision support system using AI techniques for direct decision support in acute ischemic stroke cases in adult patients. We (1) describe data and outcomes used in those systems, (2) estimate the systems' benefits compared with traditional stroke diagnosis and treatment, and (3) reported concordance with reporting standards for AI in healthcare. RESULTS: One hundred twenty-one studies met our inclusion criteria. Sixty-five were included for full extraction. In our sample, utilized data sources, methods, and reporting practices were highly heterogeneous. CONCLUSIONS: Our results suggest significant validity threats, dissonance in reporting practices, and challenges to clinical translation. We outline practical recommendations for the successful implementation of AI research in acute ischemic stroke treatment and diagnosis.

Machine Learning-Based Identification of Target Groups for Thrombectomy in Acute Stroke.

Whether endovascular thrombectomy (EVT) improves functional outcome in patients with large-vessel occlusion (LVO) stroke that do not comply with inclusion criteria of randomized controlled trials (RCTs) but that are considered for EVT in clinical practice is uncertain. We aimed to systematically identify patients with LVO stroke underrepresented in RCTs who might benefit from EVT. Following the premises that (i) patients without reperfusion after EVT represent a non-treated control group and (ii) the level of reperfusion affects outcome in patients with benefit from EVT but not in patients without treatment benefit, we systematically assessed the importance of reperfusion level on functional outcome prediction using machine learning in patients with LVO stroke treated with EVT in clinical practice (N = 5235, German-Stroke-Registry) and in patients treated with EVT or best medical management from RCTs (N = 1488, Virtual-International-Stroke-Trials-Archive). The importance of reperfusion level on outcome prediction in an RCT-like real-world cohort equaled the importance of EVT treatment allocation for outcome prediction in RCT data and was higher compared to an unselected real-world population. The importance of reperfusion level was magnified in patient groups underrepresented in RCTs, including patients with lower NIHSS scores (0-10), M2 occlusions, and lower ASPECTS (0-5 and 6-8). Reperfusion level was equally important in patients with vertebrobasilar as with anterior LVO stroke. The importance of reperfusion level for outcome prediction identifies patient target groups who likely benefit from EVT, including vertebrobasilar stroke patients and among patients underrepresented in RCT patients with low NIHSS scores, low ASPECTS, and M2 occlusions.

Anatomical labeling of intracranial arteries with deep learning in patients with cerebrovascular disease.

Brain arteries are routinely imaged in the clinical setting by various modalities, e.g., time-of-flight magnetic resonance angiography (TOF-MRA). These imaging techniques have great potential for the diagnosis of cerebrovascular disease, disease progression, and response to treatment. Currently, however, only qualitative assessment is implemented in clinical applications, relying on visual inspection. While manual or semi-automated approaches for quantification exist, such solutions are impractical in the clinical setting as they are time-consuming, involve too many processing steps, and/or neglect image intensity information. In this study, we present a deep learning-based solution for the anatomical labeling of intracranial arteries that utilizes complete information from 3D TOF-MRA images. We adapted and trained a state-of-the-art multi-scale Unet architecture using imaging data of 242 patients with cerebrovascular disease to distinguish 24 arterial segments. The proposed model utilizes vessel-specific information as well as raw image intensity information, and can thus take tissue characteristics into account. Our method yielded a performance of 0.89 macro F1 and 0.90 balanced class accuracy (bAcc) in labeling aggregated segments and 0.80 macro F1 and 0.83 bAcc in labeling detailed arterial segments on average. In particular, a higher F1 score than 0.75 for most arteries of clinical interest for cerebrovascular disease was achieved, with higher than 0.90 F1 scores in the larger, main arteries. Due to minimal pre-processing, simple usability, and fast predictions, our method could be highly applicable in the clinical setting.

Comparing Poor and Favorable Outcome Prediction With Machine Learning After Mechanical Thrombectomy in Acute Ischemic Stroke.

Background and Purpose: Outcome prediction after mechanical thrombectomy (MT) in patients with acute ischemic stroke (AIS) and large vessel occlusion (LVO) is commonly performed by focusing on favorable outcome (modified Rankin Scale, mRS 0-2) after 3 months but poor outcome representing severe disability and mortality (mRS 5 and 6) might be of equal importance for clinical decision-making. Methods: We retrospectively analyzed patients with AIS and LVO undergoing MT from 2009 to 2018. Prognostic variables were grouped in baseline clinical (A), MRI-derived variables including mismatch [apparent diffusion coefficient (ADC) and time-to-maximum (Tmax) lesion volume] (B), and variables reflecting speed and extent of reperfusion (C) [modified treatment in cerebral ischemia (mTICI) score and time from onset to mTICI]. Three different scenarios were analyzed: (1) baseline clinical parameters only, (2) baseline clinical and MRI-derived parameters, and (3) all baseline clinical, imaging-derived, and reperfusion-associated parameters. For each scenario, we assessed prediction for favorable and poor outcome with seven different machine learning algorithms. Results: In 210 patients, prediction of favorable outcome was improved after including speed and extent of recanalization [highest area under the curve (AUC) 0.73] compared to using baseline clinical variables only (highest AUC 0.67). Prediction of poor outcome remained stable by using baseline clinical variables only (highest AUC 0.71) and did not improve further by additional variables. Prediction of favorable and poor outcomes was not improved by adding MR-mismatch variables. Most important baseline clinical variables for both outcomes were age, National Institutes of Health Stroke Scale, and premorbid mRS. Conclusions: Our results suggest that a prediction of poor outcome after AIS and MT could be made based on clinical baseline variables only. Speed and extent of MT did improve prediction for a favorable outcome but is not relevant for poor outcome. An MR mismatch with small ischemic core and larger penumbral tissue showed no predictive importance.

Toward Sharing Brain Images: Differentially Private TOF-MRA Images With Segmentation Labels Using Generative Adversarial Networks.

Sharing labeled data is crucial to acquire large datasets for various Deep Learning applications. In medical imaging, this is often not feasible due to privacy regulations. Whereas anonymization would be a solution, standard techniques have been shown to be partially reversible. Here, synthetic data using a Generative Adversarial Network (GAN) with differential privacy guarantees could be a solution to ensure the patient's privacy while maintaining the predictive properties of the data. In this study, we implemented a Wasserstein GAN (WGAN) with and without differential privacy guarantees to generate privacy-preserving labeled Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) image patches for brain vessel segmentation. The synthesized image-label pairs were used to train a U-net which was evaluated in terms of the segmentation performance on real patient images from two different datasets. Additionally, the Fréchet Inception Distance (FID) was calculated between the generated images and the real images to assess their similarity. During the evaluation using the U-Net and the FID, we explored the effect of different levels of privacy which was represented by the parameter ϵ. With stricter privacy guarantees, the segmentation performance and the similarity to the real patient images in terms of FID decreased. Our best segmentation model, trained on synthetic and private data, achieved a Dice Similarity Coefficient (DSC) of 0.75 for ϵ = 7.4 compared to 0.84 for ϵ = ∞ in a brain vessel segmentation paradigm (DSC of 0.69 and 0.88 on the second test set, respectively). We identified a threshold of ϵ <5 for which the performance (DSC <0.61) became unstable and not usable. Our synthesized labeled TOF-MRA images with strict privacy guarantees retained predictive properties necessary for segmenting the brain vessels. Although further research is warranted regarding generalizability to other imaging modalities and performance improvement, our results mark an encouraging first step for privacy-preserving data sharing in medical imaging.

Combination of Radiological and Clinical Baseline Data for Outcome Prediction of Patients With an Acute Ischemic Stroke.

Background: Accurate prediction of clinical outcome is of utmost importance for choices regarding the endovascular treatment (EVT) of acute stroke. Recent studies on the prediction modeling for stroke focused mostly on clinical characteristics and radiological scores available at baseline. Radiological images are composed of millions of voxels, and a lot of information can be lost when representing this information by a single value. Therefore, in this study we aimed at developing prediction models that take into account the whole imaging data combined with clinical data available at baseline. Methods: We included 3,279 patients from the MR CLEAN Registry; a prospective, observational, multicenter registry of patients with ischemic stroke treated with EVT. We developed two approaches to combine the imaging data with the clinical data. The first approach was based on radiomics features, extracted from 70 atlas regions combined with the clinical data to train machine learning models. For the second approach, we trained 3D deep learning models using the whole images and the clinical data. Models trained with the clinical data only were compared with models trained with the combination of clinical and image data. Finally, we explored feature importance plots for the best models and identified many known variables and image features/brain regions that were relevant in the model decision process. Results: From 3,279 patients included, 1,241 (37%) patients had a good functional outcome [modified Rankin Scale (mRS) ≤ 2] and 1,954 (60%) patients had good reperfusion [modified Thrombolysis in Cerebral Infarction (eTICI) ≥ 2b]. There was no significant improvement by combining the image data to the clinical data for mRS prediction [mean area under the receiver operating characteristic (ROC) curve (AUC) of 0.81 vs. 0.80] above using the clinical data only, regardless of the approach used. Regarding predicting reperfusion, there was a significant improvement when image and clinical features were combined (mean AUC of 0.54 vs. 0.61), with the highest AUC obtained by the deep learning approach. Conclusions: The combination of radiomics and deep learning image features with clinical data significantly improved the prediction of good reperfusion. The visualization of prediction feature importance showed both known and novel clinical and imaging features with predictive values.

Generating 3D TOF-MRA volumes and segmentation labels using generative adversarial networks.

Deep learning requires large labeled datasets that are difficult to gather in medical imaging due to data privacy issues and time-consuming manual labeling. Generative Adversarial Networks (GANs) can alleviate these challenges enabling synthesis of shareable data. While 2D GANs have been used to generate 2D images with their corresponding labels, they cannot capture the volumetric information of 3D medical imaging. 3D GANs are more suitable for this and have been used to generate 3D volumes but not their corresponding labels. One reason might be that synthesizing 3D volumes is challenging owing to computational limitations. In this work, we present 3D GANs for the generation of 3D medical image volumes with corresponding labels applying mixed precision to alleviate computational constraints. We generated 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) patches with their corresponding brain blood vessel segmentation labels. We used four variants of 3D Wasserstein GAN (WGAN) with: 1) gradient penalty (GP), 2) GP with spectral normalization (SN), 3) SN with mixed precision (SN-MP), and 4) SN-MP with double filters per layer (c-SN-MP). The generated patches were quantitatively evaluated using the Fréchet Inception Distance (FID) and Precision and Recall of Distributions (PRD). Further, 3D U-Nets were trained with patch-label pairs from different WGAN models and their performance was compared to the performance of a benchmark U-Net trained on real data. The segmentation performance of all U-Net models was assessed using Dice Similarity Coefficient (DSC) and balanced Average Hausdorff Distance (bAVD) for a) all vessels, and b) intracranial vessels only. Our results show that patches generated with WGAN models using mixed precision (SN-MP and c-SN-MP) yielded the lowest FID scores and the best PRD curves. Among the 3D U-Nets trained with synthetic patch-label pairs, c-SN-MP pairs achieved the highest DSC (0.841) and lowest bAVD (0.508) compared to the benchmark U-Net trained on real data (DSC 0.901; bAVD 0.294) for intracranial vessels. In conclusion, our solution generates realistic 3D TOF-MRA patches and labels for brain vessel segmentation. We demonstrate the benefit of using mixed precision for computational efficiency resulting in the best-performing GAN-architecture. Our work paves the way towards sharing of labeled 3D medical data which would increase generalizability of deep learning models for clinical use.

Image-to-image generative adversarial networks for synthesizing perfusion parameter maps from DSC-MR images in cerebrovascular disease.

Stroke is a major cause of death or disability. As imaging-based patient stratification improves acute stroke therapy, dynamic susceptibility contrast magnetic resonance imaging (DSC-MRI) is of major interest in image brain perfusion. However, expert-level perfusion maps require a manual or semi-manual post-processing by a medical expert making the procedure time-consuming and less-standardized. Modern machine learning methods such as generative adversarial networks (GANs) have the potential to automate the perfusion map generation on an expert level without manual validation. We propose a modified pix2pix GAN with a temporal component (temp-pix2pix-GAN) that generates perfusion maps in an end-to-end fashion. We train our model on perfusion maps infused with expert knowledge to encode it into the GANs. The performance was trained and evaluated using the structural similarity index measure (SSIM) on two datasets including patients with acute stroke and the steno-occlusive disease. Our temp-pix2pix architecture showed high performance on the acute stroke dataset for all perfusion maps (mean SSIM 0.92-0.99) and good performance on data including patients with the steno-occlusive disease (mean SSIM 0.84-0.99). While clinical validation is still necessary for future studies, our results mark an important step toward automated expert-level perfusion maps and thus fast patient stratification.

An evaluation of performance measures for arterial brain vessel segmentation.

BACKGROUND: Arterial brain vessel segmentation allows utilising clinically relevant information contained within the cerebral vascular tree. Currently, however, no standardised performance measure is available to evaluate the quality of cerebral vessel segmentations. Thus, we developed a performance measure selection framework based on manual visual scoring of simulated segmentation variations to find the most suitable measure for cerebral vessel segmentation. METHODS: To simulate segmentation variations, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation. In 10 patients, we generated a set of approximately 300 simulated segmentation variations for each ground truth image. Each segmentation was visually scored based on a predefined scoring system and segmentations were ranked based on 22 performance measures common in the literature. The correlation of visual scores with performance measure rankings was calculated using the Spearman correlation coefficient. RESULTS: The distance-based performance measures balanced average Hausdorff distance (rank = 1) and average Hausdorff distance (rank = 2) provided the segmentation rankings with the highest average correlation with manual rankings. They were followed by overlap-based measures such as Dice coefficient (rank = 7), a standard performance measure in medical image segmentation. CONCLUSIONS: Average Hausdorff distance-based measures should be used as a standard performance measure in evaluating cerebral vessel segmentation quality. They can identify more relevant segmentation errors, especially in high-quality segmentations. Our findings have the potential to accelerate the validation and development of novel vessel segmentation approaches.

Synthesizing anonymized and labeled TOF-MRA patches for brain vessel segmentation using generative adversarial networks.

Anonymization and data sharing are crucial for privacy protection and acquisition of large datasets for medical image analysis. This is a big challenge, especially for neuroimaging. Here, the brain's unique structure allows for re-identification and thus requires non-conventional anonymization. Generative adversarial networks (GANs) have the potential to provide anonymous images while preserving predictive properties. Analyzing brain vessel segmentation, we trained 3 GANs on time-of-flight (TOF) magnetic resonance angiography (MRA) patches for image-label generation: 1) Deep convolutional GAN, 2) Wasserstein-GAN with gradient penalty (WGAN-GP) and 3) WGAN-GP with spectral normalization (WGAN-GP-SN). The generated image-labels from each GAN were used to train a U-net for segmentation and tested on real data. Moreover, we applied our synthetic patches using transfer learning on a second dataset. For an increasing number of up to 15 patients we evaluated the model performance on real data with and without pre-training. The performance for all models was assessed by the Dice Similarity Coefficient (DSC) and the 95th percentile of the Hausdorff Distance (95HD). Comparing the 3 GANs, the U-net trained on synthetic data generated by the WGAN-GP-SN showed the highest performance to predict vessels (DSC/95HD 0.85/30.00) benchmarked by the U-net trained on real data (0.89/26.57). The transfer learning approach showed superior performance for the same GAN compared to no pre-training, especially for one patient only (0.91/24.66 vs. 0.84/27.36). In this work, synthetic image-label pairs retained generalizable information and showed good performance for vessel segmentation. Besides, we showed that synthetic patches can be used in a transfer learning approach with independent data. This paves the way to overcome the challenges of scarce data and anonymization in medical imaging.

Outcome prediction in aneurysmal subarachnoid hemorrhage: a comparison of machine learning methods and established clinico-radiological scores.

Reliable prediction of outcomes of aneurysmal subarachnoid hemorrhage (aSAH) based on factors available at patient admission may support responsible allocation of resources as well as treatment decisions. Radiographic and clinical scoring systems may help clinicians estimate disease severity, but their predictive value is limited, especially in devising treatment strategies. In this study, we aimed to examine whether a machine learning (ML) approach using variables available on admission may improve outcome prediction in aSAH compared to established scoring systems. Combined clinical and radiographic features as well as standard scores (Hunt & Hess, WFNS, BNI, Fisher, and VASOGRADE) available on patient admission were analyzed using a consecutive single-center database of patients that presented with aSAH (n = 388). Different ML models (seven algorithms including three types of traditional generalized linear models, as well as a tree bosting algorithm, a support vector machine classifier (SVMC), a Naive Bayes (NB) classifier, and a multilayer perceptron (MLP) artificial neural net) were trained for single features, scores, and combined features with a random split into training and test sets (4:1 ratio), ten-fold cross-validation, and 50 shuffles. For combined features, feature importance was calculated. There was no difference in performance between traditional and other ML applications using traditional clinico-radiographic features. Also, no relevant difference was identified between a combined set of clinico-radiological features available on admission (highest AUC 0.78, tree boosting) and the best performing clinical score GCS (highest AUC 0.76, tree boosting). GCS and age were the most important variables for the feature combination. In this cohort of patients with aSAH, the performance of functional outcome prediction by machine learning techniques was comparable to traditional methods and established clinical scores. Future work is necessary to examine input variables other than traditional clinico-radiographic features and to evaluate whether a higher performance for outcome prediction in aSAH can be achieved.

On the usage of average Hausdorff distance for segmentation performance assessment: hidden error when used for ranking.

Average Hausdorff distance is a widely used performance measure to calculate the distance between two point sets. In medical image segmentation, it is used to compare ground truth images with segmentations allowing their ranking. We identified, however, ranking errors of average Hausdorff distance making it less suitable for applications in segmentation performance assessment. To mitigate this error, we present a modified calculation of this performance measure that we have coined "balanced average Hausdorff distance". To simulate segmentations for ranking, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation as our use-case. Adding the created errors consecutively and randomly to the ground truth, we created sets of simulated segmentations with increasing number of errors. Each set of simulated segmentations was ranked using both performance measures. We calculated the Kendall rank correlation coefficient between the segmentation ranking and the number of errors in each simulated segmentation. The rankings produced by balanced average Hausdorff distance had a significantly higher median correlation (1.00) than those by average Hausdorff distance (0.89). In 200 total rankings, the former misranked 52 whilst the latter misranked 179 segmentations. Balanced average Hausdorff distance is more suitable for rankings and quality assessment of segmentations than average Hausdorff distance.

BRAVE-NET: Fully Automated Arterial Brain Vessel Segmentation in Patients With Cerebrovascular Disease.

Introduction: Arterial brain vessel assessment is crucial for the diagnostic process in patients with cerebrovascular disease. Non-invasive neuroimaging techniques, such as time-of-flight (TOF) magnetic resonance angiography (MRA) imaging are applied in the clinical routine to depict arteries. They are, however, only visually assessed. Fully automated vessel segmentation integrated into the clinical routine could facilitate the time-critical diagnosis of vessel abnormalities and might facilitate the identification of valuable biomarkers for cerebrovascular events. In the present work, we developed and validated a new deep learning model for vessel segmentation, coined BRAVE-NET, on a large aggregated dataset of patients with cerebrovascular diseases. Methods: BRAVE-NET is a multiscale 3-D convolutional neural network (CNN) model developed on a dataset of 264 patients from three different studies enrolling patients with cerebrovascular diseases. A context path, dually capturing high- and low-resolution volumes, and deep supervision were implemented. The BRAVE-NET model was compared to a baseline Unet model and variants with only context paths and deep supervision, respectively. The models were developed and validated using high-quality manual labels as ground truth. Next to precision and recall, the performance was assessed quantitatively by Dice coefficient (DSC); average Hausdorff distance (AVD); 95-percentile Hausdorff distance (95HD); and via visual qualitative rating. Results: The BRAVE-NET performance surpassed the other models for arterial brain vessel segmentation with a DSC = 0.931, AVD = 0.165, and 95HD = 29.153. The BRAVE-NET model was also the most resistant toward false labelings as revealed by the visual analysis. The performance improvement is primarily attributed to the integration of the multiscaling context path into the 3-D Unet and to a lesser extent to the deep supervision architectural component. Discussion: We present a new state-of-the-art of arterial brain vessel segmentation tailored to cerebrovascular pathology. We provide an extensive experimental validation of the model using a large aggregated dataset encompassing a large variability of cerebrovascular disease and an external set of healthy volunteers. The framework provides the technological foundation for improving the clinical workflow and can serve as a biomarker extraction tool in cerebrovascular diseases.