Validated automatic brain extraction of head CT images

Automated segmentation of ventricular volumes and subarachnoid hemorrhage from computed tomography images: Evaluation of a rule-based pipeline approach

Changes in ventricular size, related to brain edema and hydrocephalus, as well as the extent of hemorrhage are associated with adverse outcomes in patients with subarachnoid hemorrhage (SAH). Frequently, these are measured manually using consecutive non-contrast computed tomography scans. Here, we developed a rule-based approach which incorporates both intensity and spatial normalization and utilizes user-defined thresholds and anatomical templates to segment both lateral ventricle (LV) and SAH blood volumes automatically from CT images. The algorithmic segmentations were evaluated against two expert neuroradiologists on representative slices from 20 admission scans from aneurysmal SAH patients. Previous methods have been developed to automate this time-consuming task, but they lack user feedback and are hard to implement due to large-scale data and complex design processes. Our results using automatic ventricular segmentation aligned well with expert reviewers with a median Dice coefficient of 0.81, AUC of 0.91, sensitivity of 81%, and precision of 84%. Automatic segmentation of SAH blood was most reliable near the base of the brain with a median Dice coefficient of 0.51, an AUC of 0.75, precision of 68%, and sensitivity of 50%. Ultimately, we developed a rule-based method that is easily adaptable through user feedback, generates spatially normalized segmentations that are comparable regardless of brain morphology or acquisition conditions, and automatically segments LV with good overall reliability and basal SAH blood with good precision. Our approach could benefit longitudinal studies in patients with SAH by streamlining assessment of edema and hydrocephalus progression, as well as blood resorption.

3D Vascular Segmentation Supervised by 2D Annotation of Maximum Intensity Projection

Vascular structure segmentation plays a crucial role in medical analysis and clinical applications. The practical adoption of fully supervised segmentation models is impeded by the intricacy and time-consuming nature of annotating vessels in the 3D space. This has spurred the exploration of weakly-supervised approaches that reduce reliance on expensive segmentation annotations. Despite this, existing weakly supervised methods employed in organ segmentation, which encompass points, bounding boxes, or graffiti, have exhibited suboptimal performance when handling sparse vascular structure. To alleviate this issue, we employ maximum intensity projection (MIP) to decrease the dimensionality of 3D volume to 2D image for efficient annotation, and the 2D labels are utilized to provide guidance and oversight for training 3D vessel segmentation model. Initially, we generate pseudo-labels for 3D blood vessels using the annotations of 2D projections. Subsequently, taking into account the acquisition method of the 2D labels, we introduce a weakly-supervised network that fuses 2D-3D deep features via MIP to further improve segmentation performance. Furthermore, we integrate confidence learning and uncertainty estimation to refine the generated pseudo-labels, followed by fine-tuning the segmentation network. Our method is validated on five datasets (including cerebral vessel, aorta and coronary artery), demonstrating highly competitive performance in segmenting vessels and the potential to significantly reduce the time and effort required for vessel annotation. Our code is available at: https://github.com/gzq17/Weakly-Supervised-by-MIP.

SCANED: Siamese collateral assessment network for evaluation of collaterals from ischemic damage

This study conducts collateral evaluation from ischemic damage using a deep learning-based Siamese network, addressing the challenges associated with a small and imbalanced dataset. The collateral network provides an alternative oxygen and nutrient supply pathway in ischemic stroke cases, influencing treatment decisions. Research in this area focuses on automated collateral assessment using deep learning (DL) methods to expedite decision-making processes and enhance accuracy. Our study employed a 3D ResNet-based Siamese network, referred to as SCANED, to classify collaterals as good/intermediate or poor. Utilizing non-contrast computed tomography (NCCT) images, the network automates collateral identification and assessment by analyzing tissue degeneration around the ischemic site. Relevant features from the left/right hemispheres were extracted, and Euclidean Distance (ED) was employed for similarity measurement. Finally, dichotomized classification of good/intermediate or poor collateral is performed by SCANED using an optimal threshold derived from ROC analysis. SCANED provides a sensitivity of 0.88, a specificity of 0.63, and a weighted F1 score of 0.86 in the dichotomized classification.

Adaptive mask-based brain extraction method for head CT images

Brain extraction is an important prerequisite for the automated diagnosis of intracranial lesions and determines, to a certain extent, the accuracy of subsequent lesion identification, localization, and segmentation. To address the problem that the current traditional image segmentation methods are fast in extraction but poor in robustness, while the Full Convolutional Neural Network (FCN) is robust and accurate but relatively slow in extraction, this paper proposes an adaptive mask-based brain extraction method, namely AMBBEM, to achieve brain extraction better. The method first uses threshold segmentation, median filtering, and closed operations for segmentation, generates a mask for the first time, then combines the ResNet50 model, region growing algorithm, and image properties analysis to further segment the mask, and finally complete brain extraction by multiplying the original image and the mask. The algorithm was tested on 22 test sets containing different lesions, and the results showed MPA = 0.9963, MIoU = 0.9924, and MBF = 0.9914, which were equivalent to the extraction effect of the Deeplabv3+ model. However, the method can complete brain extraction of approximately 6.16 head CT images in 1 second, much faster than Deeplabv3+, U-net, and SegNet models. In summary, this method can achieve accurate brain extraction from head CT images more quickly, creating good conditions for subsequent brain volume measurement and feature extraction of intracranial lesions.

Evolution of Severe Closed Head Injury: Assessing Ventricular Volume and Behavioral Measures at 30 and 90 Days Post-Injury

Background: Assessing functional outcomes in Severe Closed Head Injury (SCHI) is complex due to brain parenchymal changes. This study examines the Ventricles to Intracranial Volume Ratio (VBR) as a metric for these changes and its correlation with behavioral scales. Methods: Thirty-one SCHI patients were included. VBR was derived from CT scans at 3, 30, and 90 days post-injury and compared with Levels of Cognitive Functioning (LCF), Disability Rating Scale (DRS), and Early Rehabilitation Barthel Index (ERBI) assessments at 30 and 90 days. Results: Ten patients were excluded post-decompressive craniectomy or ventriculoperitoneal shunt. Findings indicated a VBR decrease at 3 days, suggesting acute phase compression, followed by an increase from 30 to 90 days, indicative of post-acute brain atrophy. VBR correlated positively with the Marshall score in the initial 72 h, positioning it as an early indicator of subsequent brain atrophy. Nevertheless, in contrast to the Marshall score, VBR had stronger associations with DRS and ERBI at 90 days. Conclusions: VBR, alongside behavioral assessments, presents a robust framework for evaluating SCHI progression. It supports early functional outcome correlations informing therapeutic approaches. VBR's reliability underscores its utility in neurorehabilitation for ongoing SCHI assessment and aiding clinical decisions.

Deep Learning-Based Extraction of Biomarkers for the Prediction of the Functional Outcome of Ischemic Stroke Patients

Accurately predicting functional outcomes in stroke patients remains challenging yet clinically relevant. While brain CTs provide prognostic information, their practical value for outcome prediction is unclear. We analyzed a multi-center cohort of 743 ischemic stroke patients (<72 h onset), including their admission brain NCCT and CTA scans as well as their clinical data. Our goal was to predict the patients' future functional outcome, measured by the 3-month post-stroke modified Rankin Scale (mRS), dichotomized into good (mRS ≤ 2) and poor (mRS > 2). To this end, we developed deep learning models to predict the outcome from CT data only, and models that incorporate other patient variables. Three deep learning architectures were tested in the image-only prediction, achieving 0.779 ± 0.005 AUC. In addition, we created a model fusing imaging and tabular data by feeding the output of a deep learning model trained to detect occlusions on CT angiograms into our prediction framework, which achieved an AUC of 0.806 ± 0.082. These findings highlight how further refinement of prognostic models incorporating both image biomarkers and clinical data could enable more accurate outcome prediction for ischemic stroke patients.

Latent Transformer Models for out-of-distribution detection

Any clinically-deployed image-processing pipeline must be robust to the full range of inputs it may be presented with. One popular approach to this challenge is to develop predictive models that can provide a measure of their uncertainty. Another approach is to use generative modelling to quantify the likelihood of inputs. Inputs with a low enough likelihood are deemed to be out-of-distribution and are not presented to the downstream predictive model. In this work, we evaluate several approaches to segmentation with uncertainty for the task of segmenting bleeds in 3D CT of the head. We show that these models can fail catastrophically when operating in the far out-of-distribution domain, often providing predictions that are both highly confident and wrong. We propose to instead perform out-of-distribution detection using the Latent Transformer Model: a VQ-GAN is used to provide a highly compressed latent representation of the input volume, and a transformer is then used to estimate the likelihood of this compressed representation of the input. We demonstrate this approach can identify images that are both far- and near- out-of-distribution, as well as provide spatial maps that highlight the regions considered to be out-of-distribution. Furthermore, we find a strong relationship between an image's likelihood and the quality of a model's segmentation on it, demonstrating that this approach is viable for filtering out unsuitable images.

A fast and accurate brain extraction method for CT head images

BACKGROUND

Brain extraction is an essential prerequisite for the automated diagnosis of intracranial lesions and determines, to a certain extent, the accuracy of subsequent lesion recognition, location, and segmentation. Segmentation using a fully convolutional neural network (FCN) yields high accuracy but a relatively slow extraction speed.

METHODS

This paper proposes an integrated algorithm, FABEM, to address the above issues. This method first uses threshold segmentation, closed operation, convolutional neural network (CNN), and image filling to generate a specific mask. Then, it detects the number of connected regions of the mask. If the number of connected regions equals 1, the extraction is done by directly multiplying with the original image. Otherwise, the mask was further segmented using the region growth method for original images with single-region brain distribution. Conversely, for images with multi-region brain distribution, Deeplabv3 + is used to adjust the mask. Finally, the mask is multiplied with the original image to complete the extraction.

RESULTS

The algorithm and 5 FCN models were tested on 24 datasets containing different lesions, and the algorithm's performance showed MPA = 0.9968, MIoU = 0.9936, and MBF = 0.9963, comparable to the Deeplabv3+. Still, its extraction speed is much faster than the Deeplabv3+. It can complete the brain extraction of a head CT image in about 0.43 s, about 3.8 times that of the Deeplabv3+.

CONCLUSION

Thus, this method can achieve accurate brain extraction from head CT images faster, creating a good basis for subsequent brain volume measurement and feature extraction of intracranial lesions.

Imaging biomarkers of cerebral edema automatically extracted from routine CT scans of large vessel occlusion strokes

BACKGROUND AND PURPOSE

Volumetric and densitometric biomarkers have been proposed to better quantify cerebral edema after stroke, but their relative performance has not been rigorously evaluated.

METHODS

Patients with large vessel occlusion stroke from three institutions were analyzed. An automated pipeline extracted brain, cerebrospinal fluid (CSF), and infarct volumes from serial CTs. Several biomarkers were measured: change in global CSF volume from baseline (ΔCSF); ratio of CSF volumes between hemispheres (CSF ratio); and relative density of infarct region compared with mirrored contralateral region (net water uptake [NWU]). These were compared to radiographic standards, midline shift and relative hemispheric volume (RHV) and malignant edema, defined as deterioration resulting in need for osmotic therapy, decompressive surgery, or death.

RESULTS

We analyzed 255 patients with 210 baseline CTs, 255 24-hour CTs, and 81 72-hour CTs. Of these, 35 (14%) developed malignant edema and 63 (27%) midline shift. CSF metrics could be calculated for 310 (92%), while NWU could only be obtained from 193 (57%). Peak midline shift was correlated with baseline CSF ratio (ρ = -.22) and with CSF ratio and ΔCSF at 24 hours (ρ = -.55/.63) and 72 hours (ρ = -.66/.69), but not with NWU (ρ = .15/.25). Similarly, CSF ratio was correlated with RHV (ρ = -.69/-.78), while NWU was not. Adjusting for age, National Institutes of Health Stroke Scale, tissue plasminogen activator treatment, and Alberta Stroke Program Early CT Score, CSF ratio (odds ratio [OR]: 1.95 per 0.1, 95% confidence interval [CI]: 1.52-2.59) and ΔCSF at 24 hours (OR: 1.87 per 10%, 95% CI: 1.47-2.49) were associated with malignant edema.

CONCLUSION

CSF volumetric biomarkers can be automatically measured from almost all routine CTs and correlate better with standard edema endpoints than net water uptake.

Predicting vasospasm risk using first presentation aneurysmal subarachnoid hemorrhage volume: A semi-automated CT image segmentation analysis using ITK-SNAP

PURPOSE

Cerebral vasospasm following aneurysmal subarachnoid hemorrhage (aSAH) is a significant complication associated with poor neurological outcomes. We present a novel, semi-automated pipeline, implemented in the open-source medical imaging analysis software ITK-SNAP, to segment subarachnoid blood volume from initial CT head (CTH) scans and use this to predict future radiological vasospasm.

METHODS

42 patients were admitted between February 2020 and December 2021 to our tertiary neurosciences center, and whose initial referral CTH scan was used for this retrospective cohort study. Blood load was segmented using a semi-automated random forest classifier and active contour evolution implemented in ITK-SNAP. Clinical data were extracted from electronic healthcare records in order to fit models aimed at predicting radiological vasospasm risk.

RESULTS

Semi-automated segmentations demonstrated excellent agreement with manual, expert-derived volumes (mean Dice coefficient = 0.92). Total normalized blood volume, extracted from CTH images at first presentation, was significantly associated with greater odds of later radiological vasospasm, increasing by approximately 7% for each additional cm3 of blood (OR = 1.069, 95% CI: 1.021-1.120; p < .005). Greater blood volume was also significantly associated with vasospasm of a higher Lindegaard ratio, of longer duration, and a greater number of discrete episodes. Total blood volume predicted radiological vasospasm with a greater accuracy as compared to the modified Fisher scale (AUC = 0.86 vs 0.70), and was of independent predictive value.

CONCLUSION

Semi-automated methods provide a plausible pipeline for the segmentation of blood from CT head images in aSAH, and total blood volume is a robust, extendable predictor of radiological vasospasm, outperforming the modified Fisher scale. Greater subarachnoid blood volume significantly increases the odds of subsequent vasospasm, its time course and its severity.

SELF-SUPERVISED LEARNING WITH RADIOLOGY REPORTS, A COMPARATIVE ANALYSIS OF STRATEGIES FOR LARGE VESSEL OCCLUSION AND BRAIN CTA IMAGES

Scarcity of labels for medical images is a significant barrier for training representation learning approaches based on deep neural networks. This limitation is also present when using imaging data collected during routine clinical care stored in picture archiving communication systems (PACS), as these data rarely have attached the high-quality labels required for medical image computing tasks. However, medical images extracted from PACS are commonly coupled with descriptive radiology reports that contain significant information and could be leveraged to pre-train imaging models, which could serve as starting points for further task-specific fine-tuning. In this work, we perform a head-to-head comparison of three different self-supervised strategies to pre-train the same imaging model on 3D brain computed tomography angiogram (CTA) images, with large vessel occlusion (LVO) detection as the downstream task. These strategies evaluate two natural language processing (NLP) approaches, one to extract 100 explicit radiology concepts (Rad-SpatialNet) and the other to create general-purpose radiology reports embeddings (DistilBERT). In addition, we experiment with learning radiology concepts directly or by using a recent self-supervised learning approach (CLIP) that learns by ranking the distance between language and image vector embeddings. The LVO detection task was selected because it requires 3D imaging data, is clinically important, and requires the algorithm to learn outputs not explicitly stated in the radiology report. Pre-training was performed on an unlabeled dataset containing 1,542 3D CTA - reports pairs. The downstream task was tested on a labeled dataset of 402 subjects for LVO. We find that the pre-training performed with CLIP-based strategies improve the performance of the imaging model to detect LVO compared to a model trained only on the labeled data. The best performance was achieved by pre-training using the explicit radiology concepts and CLIP strategy.

Automated Detection of Lesions in Patients with Traumatic Brain Injury using Brain CT Images: Concept Note and Proposed Method

Accurate and early interpretation of CT scan images in TBI patients reduces the critical time for diagnosis and management. As mentioned in other studies, automated CT interpretation using the feature extraction method is a rapid and accurate tool. Despite several studies on the machine and deep learning employing algorithms for automated CT interpretations, it has its challenges. This study presents a concept note and proposes a feature-based computer-aided diagnostic method to perform automated CT interpretation in TBI. The method consists of preprocessing, segmentation, and extraction. We have described a simple way of classifying the CT scan head into five circumferential zones in this method. The zones are identified quickly based on the anatomic characteristics and specific pathologies that affect each zone. Then, we have provided an overview of different pathologies affecting each of these zones. Utilizing these zones for automated CT interpretation will also be a helpful resource for concerned physicians during the odd and rush hours.

Lesion-preserving unpaired image-to-image translation between MRI and CT from ischemic stroke patients

PURPOSE

Multiple medical imaging modalities are used for clinical follow-up ischemic stroke analysis. Mixed-modality datasets are challenging, both for clinical rating purposes and for training machine learning models. While image-to-image translation methods have been applied to harmonize stroke patient images to a single modality, they have only been used for paired data so far. In the more common unpaired scenario, the standard cycle-consistent generative adversarial network (CycleGAN) method is not able to translate the stroke lesions properly. Thus, the aim of this work was to develop and evaluate a novel image-to-image translation regularization approach for unpaired 3D follow-up stroke patient datasets.

METHODS

A modified CycleGAN was used to translate images between 238 non-contrast computed tomography (NCCT) and 244 fluid-attenuated inversion recovery (FLAIR) MRI datasets, two of the most relevant follow-up modalities in clinical practice. We introduced an additional attention-guided mechanism to encourage an improved translation of the lesion and a gradient-consistency loss to preserve structural brain morphology.

RESULTS

The proposed modifications were able to preserve the overall quality provided by the CycleGAN translation. This was confirmed by the FID score and gradient correlation results. Furthermore, the lesion preservation was significantly improved compared to a standard CycleGAN. This was evaluated for location and volume with segmentation models, which were trained on real datasets and applied to the translated test images. Here, the Dice score coefficient resulted in 0.81 and 0.62 for datasets translated to FLAIR and NCCT, respectively, compared to 0.57 and 0.50 for the corresponding datasets translated using a standard CycleGAN. Finally, an analysis of the distribution of mean lesion intensities showed substantial improvements.

CONCLUSION

The results of this work show that the proposed image-to-image translation method is effective at preserving stroke lesions in unpaired modality translation, supporting its potential as a tool for stroke image analysis in real-life scenarios.

Intracerebral Hemorrhage Segmentation on Noncontrast Computed Tomography Using a Masked Loss Function U-Net Approach

OBJECTIVE

Intracerebral hemorrhage (ICH) volume is a strong predictor of outcome in patients presenting with acute hemorrhagic stroke. It is necessary to segment the hematoma for ICH volume estimation and for computerized extraction of features, such as spot sign, texture parameters, or extravasated iodine content at dual-energy computed tomography. Manual and semiautomatic segmentation methods to delineate the hematoma are tedious, user dependent, and require trained personnel. This article presents a convolutional neural network to automatically delineate ICH from noncontrast computed tomography scans of the head.

METHODS

A model combining a U-Net architecture with a masked loss function was trained on standard noncontrast computed tomography images that were down sampled to 256 × 256 size. Data augmentation was applied to prevent overfitting, and the loss score was calculated using the soft Dice loss function. The Dice coefficient and the Hausdorff distance were computed to quantitatively evaluate the segmentation performance of the model, together with the sensitivity and specificity to determine the ICH detection accuracy.

RESULTS

The results demonstrate a median Dice coefficient of 75.9% and Hausdorff distance of 2.65 pixels in segmentation performance, with a detection sensitivity of 77.0% and specificity of 96.2%.

CONCLUSIONS

The proposed masked loss U-Net is accurate in the automatic segmentation of ICH. Future research should focus on increasing the detection sensitivity of the model and comparing its performance with other model architectures.

Does the Mutation Type Affect the Response to Cranial Vault Expansion in Children With Apert Syndrome?

LEVEL OF EVIDENCE

III.

Dataset on acute stroke risk stratification from CT angiographic radiomics

With advances in high-throughput image processing technologies and increasing availability of medical mega-data, the growing field of radiomics opened the door for quantitative analysis of medical images for prediction of clinically relevant information. One clinical area in which radiomics have proven useful is stroke neuroimaging, where rapid treatment triage is vital for patient outcomes and automated decision assistance tools have potential for significant clinical impact. Recent research, for example, has applied radiomics features extracted from CT angiography (CTA) images and a machine learning framework to facilitate risk-stratification in acute stroke. We here provide methodological guidelines and radiomics data supporting the referenced article "CT angiographic radiomics signature for risk-stratification in anterior large vessel occlusion stroke." The data were extracted from the stroke center registry at Yale New Haven Hospital between 1/1/2014 and 10/31/2020; and Geisinger Medical Center between 1/1/2016 and 12/31/2019. It includes detailed radiomics features of the anterior circulation territories on admission CTA scans in stroke patients with large vessel occlusion stroke who underwent thrombectomy. We also provide the methodological details of the analysis framework utilized for training, optimization, validation and external testing of the machine learning and feature selection algorithms. With the goal of advancing the feasibility and quality of radiomics-based analyses to improve patient care within and beyond the field of stroke, the provided data and methodological support can serve as a baseline for future studies applying radiomics algorithms to machine-learning frameworks, and allow for analysis and utilization of radiomics features extracted in this study.

[18F]FE-PE2I PET is a feasible alternative to [123I]FP-CIT SPECT for dopamine transporter imaging in clinically uncertain parkinsonism

BACKGROUND

Dopamine transporter (DAT) imaging of striatum is clinically used in Parkinson's disease (PD) and neurodegenerative parkinsonian syndromes (PS) especially in the early disease stages. The aim of the present study was to evaluate the diagnostic performance of the recently developed tracer for DAT imaging [¹⁸F]FE-PE2I PET/CT to the reference standard [¹²³I]FP-CIT SPECT.

METHODS

Ninety-eight unselected patients referred for DAT imaging were included prospectively and consecutively and evaluated with [¹⁸F]FE-PE2I PET/CT and [¹²³I]FP-CIT SPECT on two separate days. PET and SPECT scans were categorized independently by two blinded expert readers as either normal, vascular changes, or mixed. Semiquantitative values were obtained for each modality and compared regarding effect size using Glass' delta.

RESULTS

Fifty-six of the [¹²³I]FP-CIT SPECT scans were considered abnormal (52 caused by PS, 4 by infarctions). Using [¹⁸F]FE-PE2I PET/CT, 95 of the 98 patients were categorized identically to SPECT as PS or non-PS with a sensitivity of 0.94 [0.84-0.99] and a specificity of 1.00 [0.92-1.00]. Inter-reader agreement for [¹⁸F]FE-PE2I PET with a kappa of 0.97 [0.89-1.00] was comparable to the agreement for [¹²³I]FP-CIT SPECT of 0.96 [0.76-1.00]. Semiquantitative values for short 10-min reconstructions of [¹⁸F]FE-PE2I PET/CT were comparable to longer reconstructions. The effect size for putamen/caudate nucleus ratio was significantly increased using PET compared to SPECT.

CONCLUSIONS

The high correspondence of [¹⁸F]FE-PE2I PET compared to reference standard [¹²³I]FP-CIT SPECT establishes [¹⁸F]FE-PE2I PET as a feasible PET tracer for clinical use with favourable scan logistics.

Bayesian deep learning outperforms clinical trial estimators of intracerebral and intraventricular hemorrhage volume

BACKGROUND AND PURPOSE

Intracerebral hemorrhage (ICH) and intraventricular hemorrhage (IVH) clinical trials rely on manual linear and semi-quantitative (LSQ) estimators like the ABC/2, modified Graeb and IVH scores for timely volumetric estimation from CT. Deep learning (DL) volumetrics of ICH have recently approached the accuracy of gold-standard planimetry. However, DL and LSQ strategies have been limited by unquantified uncertainty, in particular when ICH and IVH estimates intersect. Bayesian deep learning methods can be used to approximate uncertainty, presenting an opportunity to improve quality assurance in clinical trials.

METHODS

A DL model was trained to simultaneously segment ICH and IVH using diagnostic CT data from the Minimally Invasive Surgery Plus Alteplase for ICH Evacuation (MISTIE) III and Clot Lysis: Evaluating Accelerated Resolution of IVH (CLEAR) III clinical trials. Bayesian uncertainty approximation was performed using Monte-Carlo dropout. We compared the performance of our model with estimators used in the CLEAR IVH and MISTIE II trials. The reliability of planimetry, DL, and LSQ volumetrics in the setting of high ICH and IVH intersection is quantified using consensus estimates.

RESULTS

Our DL model produced volume correlations and median Dice scores of .994 and .946 for ICH in MISTIE II, and .980 and .863 for IVH in CLEAR IVH, respectively, outperforming LSQ estimates from the clinical trials. We found significant linear relationships between ICH uncertainty, Dice scores (r = -.849), and relative volume difference (r = .735).

CONCLUSION

In our validation clinical trial dataset, DL models with Bayesian uncertainty approximation provided superior volumetric estimates to LSQ methods with real-time estimates of model uncertainty.

Machine learning-based automatic estimation of cortical atrophy using brain computed tomography images

Cortical atrophy is measured clinically according to established visual rating scales based on magnetic resonance imaging (MRI). Although brain MRI is the primary imaging marker for neurodegeneration, computed tomography (CT) is also widely used for the early detection and diagnosis of dementia. However, they are seldom investigated. Therefore, we developed a machine learning algorithm for the automatic estimation of cortical atrophy on brain CT. Brain CT images (259 Alzheimer's dementia and 55 cognitively normal subjects) were visually rated by three neurologists and used for training. We constructed an algorithm by combining the convolutional neural network and regularized logistic regression (RLR). Model performance was then compared with that of neurologists, and feature importance was measured. RLR provided fast and reliable automatic estimations of frontal atrophy (75.2% accuracy, 93.6% sensitivity, 67.2% specificity, and 0.87 area under the curve [AUC]), posterior atrophy (79.6% accuracy, 87.2% sensitivity, 75.9% specificity, and 0.88 AUC), right medial temporal atrophy (81.2% accuracy, 84.7% sensitivity, 79.6% specificity, and 0.88 AUC), and left medial temporal atrophy (77.7% accuracy, 91.1% sensitivity, 72.3% specificity, and 0.90 AUC). We concluded that RLR-based automatic estimation of brain CT provided a comprehensive rating of atrophy that can potentially support physicians in real clinical settings.

Stroke population–specific neuroanatomical CT-MRI brain atlas

Purpose

Development of a freely available stroke population–specific anatomical CT/MRI atlas with a reliable normalisation pipeline for clinical CT.

Methods

By reviewing CT scans in suspected stroke patients and filtering the AIBL MRI database, respectively, we collected 50 normal-for-age CT and MRI scans to build a standard-resolution CT template and a high-resolution MRI template. The latter was manually segmented into anatomical brain regions. We then developed and validated a MRI to CT registration pipeline to align the MRI atlas onto the CT template. Finally, we developed a CT-to-CT-normalisation pipeline and tested its reliability by calculating Dice coefficient (Dice) and Average Hausdorff Distance (AHD) for predefined areas in 100 CT scans from ischaemic stroke patients.

Results

The resulting CT/MRI templates were age and sex matched to a general stroke population (median age 71.9 years (62.1–80.2), 60% male). Specifically, this accounts for relevant structural changes related to aging, which may affect registration. Applying the validated MRI to CT alignment (Dice > 0.78, Average Hausdorff Distance < 0.59 mm) resulted in our final CT-MRI atlas. The atlas has 52 manually segmented regions and covers the whole brain. The alignment of four cortical and subcortical brain regions with our CT-normalisation pipeline was reliable for small/medium/large infarct lesions (Dice coefficient > 0.5).

Conclusion

The newly created CT-MRI brain atlas has the potential to standardise stroke lesion segmentation. Together with the automated normalisation pipeline, it allows analysis of existing and new datasets to improve prediction tools for stroke patients (free download at https://forms.office.com/r/v4t3sWfbKs).

Supplementary Information

The online version contains supplementary material available at 10.1007/s00234-021-02875-9.

A Comparison of Cranial Cavity Extraction Tools for Non-contrast Enhanced CT Scans in Acute Stroke Patients

Cranial cavity extraction is often the first step in quantitative neuroimaging analyses. However, few automated, validated extraction tools have been developed for non-contrast enhanced CT scans (NECT). The purpose of this study was to compare and contrast freely available tools in an unseen dataset of real-world clinical NECT head scans in order to assess the performance and generalisability of these tools. This study included data from a demographically representative sample of 428 patients who had completed NECT scans following hospitalisation for stroke. In a subset of the scans (n = 20), the intracranial spaces were segmented using automated tools and compared to the gold standard of manual delineation to calculate accuracy, precision, recall, and dice similarity coefficient (DSC) values. Further, three readers independently performed regional visual comparisons of the quality of the results in a larger dataset (n = 428). Three tools were found; one of these had unreliable performance so subsequent evaluation was discontinued. The remaining tools included one that was adapted from the FMRIB software library (fBET) and a convolutional neural network- based tool (rBET). Quantitative comparison showed comparable accuracy, precision, recall and DSC values (fBET: 0.984 ± 0.002; rBET: 0.984 ± 0.003; p = 0.99) between the tools; however, intracranial volume was overestimated. Visual comparisons identified characteristic regional differences in the resulting cranial cavity segmentations. Overall fBET had highest visual quality ratings and was preferred by the readers in the majority of subject results (84%). However, both tools produced high quality extractions of the intracranial space and our findings should improve confidence in these automated CT tools. Pre- and post-processing techniques may further improve these results.

Outcome Prediction of Patients with Intracerebral Hemorrhage by Measurement of Lesion Volume in the Corticospinal Tract on Computed Tomography

Objective:

This study investigated the potential utility of computed tomography for outcome prediction in patients with intracerebral hemorrhage.

Methods:

Patients with putaminal and/or thalamic hemorrhage for whom computed tomography images were acquired in our hospital emergency room soon after onset were retrospectively enrolled. Outcome measurements were obtained at discharge from the convalescent rehabilitation ward of our affiliated hospital. Hemiparesis was evaluated using the total score of the motor component of the Stroke Impairment Assessment Set (SIAS-motor; null to full, 0 to 25), the motor component of the Functional Independence Measure (FIM-motor; null to full, 13 to 91), and the total length of hospital stay. After registration of the computed tomography images to the standard brain, the volumes of the hematoma lesions located in the corticospinal tract were calculated. The correlation between the corticospinal tract lesion volumes and the outcome measurements was assessed using Spearman’s rank correlation test.

Results:

Thirty patients were entered into the final analytical database. Corticospinal tract lesion volumes ranged from 0.002 to 4.302 ml (median, 1.478). SIAS-motor scores ranged from 0 to 25 (median, 20), FIM-motor scores ranged from 15 to 91 (median, 80.5), and the total length of hospital stay ranged from 31 to 194 days (median, 106.5). All correlation tests were statistically significant (P <0.01). The strongest correlation was for SIAS-motor total (R=–0.710), followed by FIM-motor (R=–0.604) and LOS (R=0.493).

Conclusions:

These findings suggest that conventional computed tomography images may be useful for outcome prediction in patients with intracerebral hemorrhage.

Quantitative Susceptibility Mapping for Staging Acute Cerebral Hemorrhages: Comparing the Conventional and Multiecho Complex Total Field Inversion magnetic resonance imaging MR Methods

BACKGROUND

The perceived acuity of intracerebral hemorrhage (ICH) impacts the management of patients, both within emergent and outpatient/urgent settings. Morphology enabled dipole inversion (MEDI) quantitative susceptibility imaging (QSM) has improved characterization of ICH acuity, despite outstanding limitations in distinguishing blood products.

PURPOSE/HYPOTHESIS

Using improved susceptibility quantification, novel postprocessing QSM method from multiecho complex total field inversion (mcTFI) may better discriminate between acute and subacute ICH, compared to MEDI.

STUDY TYPE

Retrospective cohort study.

SUBJECTS

A total of 121 subjects enrolled following positive computerized tomography (CT) findings for ICH. Subjects were grouped based on time between admission and MR imaging: hyperacute (<24 hours), acute (1-3 days), early subacute (3-7 days), and late subacute (7-18 days).

FIELD STRENGTH/SEQUENCE

A multiecho gradient echo sequence at 3.0 T was paired with clinical noncontrast CT imaging.

ASSESSMENT

A quantitative index (CTindex ) was derived based on relative intensities of blood on noncontrast CT. All images were co-registered, from which QSM parameters within the ICH area were assessed across groups, as well as the correlation with CTindex .

STATISTICAL TESTS

Group differences were assessed using ANOVAs. Linear regressions between the CTindex , MEDI, and mcTFI measurements were used to assess their relationships. Statistical significance was set at P < 0.05.

RESULTS

A total of 21 hyperacute, 72 acute, 21 early subacute, and 7 late-subacute patients were included in this analysis. Significant changes in blood susceptibility were found over time for the MEDI and mcTFI, although mcTFI better differentiated the hyperacute/acute from subacute stages. CTindex values within the ICH were more strongly correlated with mcTFI QSM (r = 0.727) than MEDI (r = 0.412) QSM.

DATA CONCLUSION

McTFI susceptibility estimation demonstrated better correlation with ICH acuity as suggested by CT, providing an improved method to assess acuity of intracranial blood products in clinical settings to identify cases that may require acute intervention.

LEVEL OF EVIDENCE

4 TECHNICAL EFFICACY STAGE: 2.

Comparison of convolutional neural networks for detecting large vessel occlusion on computed tomography angiography

PURPOSE

Artificial intelligence diagnosis and triage of large vessel occlusion may quicken clinical response for a subset of time-sensitive acute ischemic stroke patients, improving outcomes. Differences in architectural elements within data-driven convolutional neural network (CNN) models impact performance. Foreknowledge of effective model architectural elements for domain-specific problems can narrow the search for candidate models and inform strategic model design and adaptation to optimize performance on available data. Here, we study CNN architectures with a range of learnable parameters and which span the inclusion of architectural elements, such as parallel processing branches and residual connections with varying methods of recombining residual information.

METHODS

We compare five CNNs: ResNet-50, DenseNet-121, EfficientNet-B0, PhiNet, and an Inception module-based network, on a computed tomography angiography large vessel occlusion detection task. The models were trained and preliminarily evaluated with 10-fold cross-validation on preprocessed scans (n = 240). An ablation study was performed on PhiNet due to superior cross-validated test performance across accuracy, precision, recall, specificity, and F1 score. The final evaluation of all models was performed on a withheld external validation set (n = 60) and these predictions were subsequently calibrated with sigmoid curves.

RESULTS

Uncalibrated results on the withheld external validation set show that DenseNet-121 had the best average performance on accuracy, precision, recall, specificity, and F1 score. After calibration DenseNet-121 maintained superior performance on all metrics except recall.

CONCLUSIONS

The number of learnable parameters in our five models and best-ablated PhiNet directly related to cross-validated test performance-the smaller the model the better. However, this pattern did not hold when looking at generalization on the withheld external validation set. DenseNet-121 generalized the best; we posit this was due to its heavy use of residual connections utilizing concatenation, which causes feature maps from earlier layers to be used deeper in the network, while aiding in gradient flow and regularization.

Head CT: Toward Making Full Use of the Information the X-Rays Give

Although clinical head CT images are typically interpreted qualitatively, automated methods applied to routine clinical head CTs enable quantitative assessment of brain volume, brain parenchymal fraction, brain radiodensity, and brain radiomass. These metrics gain clinical meaning when viewed relative to a reference database and expressed as quantile regression values. Quantitative imaging data can aid in objective reporting and in the identification of outliers, with possible diagnostic implications. The comparison to a reference database necessitates standardization of head CT imaging parameters and protocols. Future research is needed to learn the effects of virtual monochromatic imaging on the quantitative characteristics of head CT images.

3D dissimilar-siamese-u-net for hyperdense Middle cerebral artery sign segmentation

The hyperdense middle cerebral artery sign (HMCAS) representing a thromboembolus has been declared as a vital CT finding for intravascular thrombus in the diagnosis of acute ischemia stroke. Early recognition of HMCAS can assist in patient triage and subsequent thrombolysis or thrombectomy treatment. A total of 624 annotated head non-contrast-enhanced CT (NCCT) image scans were retrospectively collected from multiple public hospitals in Hong Kong. In this study, we present a deep Dissimilar-Siamese-U-Net (DSU-Net) that is able to precisely segment the lesions by integrating Siamese and U-Net architectures. The proposed framework consists of twin sub-networks that allow inputs of left and right hemispheres in head NCCT images separately. The proposed Dissimilar block fully explores the feature representation of the differences between the bilateral hemispheres. Ablation studies were carried out to validate the performance of various components of the proposed DSU-Net. Our findings reveal that the proposed DSU-Net provides a novel approach for HMCAS automatic segmentation and it outperforms the baseline U-Net and many state-of-the-art models for clinical practice.

Automatic segmentation, feature extraction and comparison of healthy and stroke cerebral vasculature

Accurate segmentation of cerebral vasculature and a quantitative assessment of its morphology is critical to various diagnostic and therapeutic purposes and is pertinent to studying brain health and disease. However, this is still a challenging task due to the complexity of the vascular imaging data. We propose an automated method for cerebral vascular segmentation without the need of any manual intervention as well as a method to skeletonize the binary segmented map to extract vascular geometric features and characterize vessel structure. We combine a Hessian-based probabilistic vessel-enhancing filtering with an active-contour-based technique to segment magnetic resonance and computed tomography angiograms (MRA and CTA) and subsequently extract the vessel centerlines and diameters to calculate the geometrical properties of the vasculature. Our method was validated using a 3D phantom of the Circle-of-Willis region, demonstrating 84% mean Dice similarity coefficient (DSC) and 85% mean Pearson's correlation coefficient (PCC) with minimal modified Hausdorff distance (MHD) error (3 surface pixels at most), and showed superior performance compared to existing segmentation algorithms upon quantitative comparison using DSC, PCC and MHD. We subsequently applied our algorithm to a dataset of 40 subjects, including 1) MRA scans of healthy subjects (n = 10, age = 30 ± 9), 2) MRA scans of stroke patients (n = 10, age = 51 ± 15), 3) CTA scans of healthy subjects (n = 10, age = 62 ± 12), and 4) CTA scans of stroke patients (n = 10, age = 68 ± 11), and obtained a quantitative comparison between the stroke and normal vasculature for both imaging modalities. The vascular network in stroke patients compared to age-adjusted healthy subjects was found to have a significantly (p < 0.05) higher tortuosity (3.24 ± 0.88 rad/cm vs. 7.17 ± 1.61 rad/cm for MRA, and 4.36 ± 1.32 rad/cm vs. 7.80 ± 0.92 rad/cm for CTA), higher fractal dimension (1.36 ± 0.28 vs. 1.71 ± 0.14 for MRA, and 1.56 ± 0.05 vs. 1.69 ± 0.20 for CTA), lower total length (3.46 ± 0.99 m vs. 2.20 ± 0.67 m for CTA), lower total volume (61.80 ± 18.79 ml vs. 34.43 ± 22.9 ml for CTA), lower average diameter (2.4 ± 0.21 mm vs. 2.18 ± 0.07 mm for CTA), and lower average branch length (4.81 ± 1.97 mm vs. 8.68 ± 2.03 mm for MRA), respectively. We additionally studied the change in vascular features with respect to aging and imaging modality. While we observed differences between features as a result of aging, statistical analysis did not show any significant differences, whereas we found that the number of branches were significantly different (p < 0.05) between the two imaging modalities (201 ± 73 for MRA vs. 189 ± 69 for CTA). Our segmentation and feature extraction algorithm can be applied on any imaging modality and can be used in the future to automatically obtain the 3D segmented vasculature for diagnosis and treatment planning as well as to study morphological changes due to stroke and other cerebrovascular diseases (CVD) in the clinic.

Pediatric Head CT: Automated Quantitative Analysis with Quantile Regression

BACKGROUND AND PURPOSE

Together with quantile regression methods, such a model would have the potential for clinical utility through automated quantitative comparison of individual cases relative to their age and gender-matched peer group. Our aim was to demonstrate the automated processing of digital clinical head CT data in the development of a clinically useful model of age-related changes of the brain in the first 2 decades of life.

MATERIALS AND METHODS

A total of 415 (209 female) consecutive, clinical head CTs with radiographically normal findings from patients from birth through 20 years of age were retrospectively selected and subjected to automated segmentation. Brain volume, brain parenchymal fraction, brain radiodensity, and brain radiomass were assessed as a function of patient age. Statistical modeling and quantile regression were performed.

RESULTS

Brain volume increased from 400 cm3 at birth to 1350 cm3 at 20 years of age (>3-fold). Males had a slightly steeper growth trajectory than females, with approximately 8% difference in volume between the sexes established in the first few years of life. Brain parenchymal fraction was variable at younger than 2 years of age, stabilizing between 0.85 and 0.92 at 2-3 years of age. Brain mean radiodensity was lower at birth (24 HU) and increased through 3 years of age, after which it stabilized near 30 HU, an approximately 25% increase. The product of brain volume and mean brain radiodensity (radiomass), increased from 700 HU × mL at birth to 3900 HU × mL, a 5.6-fold increase, with approximately 5% difference between males and females at 20 years. Quantile regression enables a given metric to be interpreted relative to an age- and sex-matched peer group.

CONCLUSIONS

Automated segmentation of clinical head CT images permitted the generation of a reference database for quantitative analysis of pediatric and adolescent brains. Quantile regression facilitates clinical application.

Multimodal Predictive Modeling of Endovascular Treatment Outcome for Acute Ischemic Stroke Using Machine-Learning

Background and Purpose:

This study assessed the predictive performance and relative importance of clinical, multimodal imaging, and angiographic characteristics for predicting the clinical outcome of endovascular treatment for acute ischemic stroke.

Methods:

A consecutive series of 246 patients with acute ischemic stroke and large vessel occlusion in the anterior circulation who underwent endovascular treatment between April 2014 and January 2018 was analyzed. Clinical, conventional imaging (electronic Alberta Stroke Program Early CT Score, acute ischemic volume, site of vessel occlusion, and collateral score), and advanced imaging characteristics (CT-perfusion with quantification of ischemic penumbra and infarct core volumes) before treatment as well as angiographic (interval groin puncture-recanalization, modified Thrombolysis in Cerebral Infarction score) and postinterventional clinical (National Institutes of Health Stroke Scale score after 24 hours) and imaging characteristics (electronic Alberta Stroke Program Early CT Score, final infarction volume after 18–36 hours) were assessed. The modified Rankin Scale (mRS) score at 90 days (mRS-90) was used to measure patient outcome (favorable outcome: mRS-90 ≤2 versus unfavorable outcome: mRS-90 >2). Machine-learning with gradient boosting classifiers was used to assess the performance and relative importance of the extracted characteristics for predicting mRS-90.

Results:

Baseline clinical and conventional imaging characteristics predicted mRS-90 with an area under the receiver operating characteristics curve of 0.740 (95% CI, 0.733–0.747) and an accuracy of 0.711 (95% CI, 0.705–0.717). Advanced imaging with CT-perfusion did not improved the predictive performance (area under the receiver operating characteristics curve, 0.747 [95% CI, 0.740–0.755]; accuracy, 0.720 [95% CI, 0.714–0.727]; P =0.150). Further inclusion of angiographic and postinterventional characteristics significantly improved the predictive performance (area under the receiver operating characteristics curve, 0.856 [95% CI, 0.850–0.861]; accuracy, 0.804 [95% CI, 0.799–0.810]; P <0.001). The most important parameters for predicting mRS 90 were National Institutes of Health Stroke Scale score after 24 hours (importance =100%), premorbid mRS score (importance =44%) and final infarction volume on postinterventional CT after 18 to 36 hours (importance =32%).

Conclusions:

Integrative assessment of clinical, multimodal imaging, and angiographic characteristics with machine-learning allowed to accurately predict the clinical outcome following endovascular treatment for acute ischemic stroke. Thereby, premorbid mRS was the most important clinical predictor for mRS-90, and the final infarction volume was the most important imaging predictor, while the extent of hemodynamic impairment on CT-perfusion before treatment had limited importance.

MP3: Medical Software for Processing Multi-Parametric Images Pipelines

This article presents an open source software able to convert, display, and process medical images. It differentiates itself from the existing software by its ability to design complex processing pipelines and to wisely execute them on a large databases. An MP3 pipeline can contain unlimited homemade or ready-made processes and can be carried out with a parallel execution system. As a viewer, MP3 allows display of up to four images together and to draw Regions Of Interest (ROI). Two applications showing the strengths of the software are presented as examples: a preclinical study involving Magnetic Resonance Imaging (MRI) data and a clinical one involving Computed Tomography (CT) images. MP3 is downloadable at https://github.com/nifm-gin/MP3.

3D Deep Neural Network Segmentation of Intracerebral Hemorrhage: Development and Validation for Clinical Trials

Intracranial hemorrhage (ICH) occurs when a blood vessel ruptures in the brain. This leads to significant morbidity and mortality, the likelihood of which is predicated on the size of the bleeding event. X-ray computed tomography (CT) scans allow clinicians and researchers to qualitatively and quantitatively diagnose hemorrhagic stroke, guide interventions and determine inclusion criteria of patients in clinical trials. There is no currently available open source, validated tool to quickly segment hemorrhage. Using an automated pipeline and 2D and 3D deep neural networks, we show that we can quickly and accurately estimate ICH volume with high agreement with time-consuming manual segmentation. The training and validation datasets include significant heterogeneity in terms of pathology, such as the presence of intraventricular (IVH) or subdural hemorrhages (SDH) as well as variable image acquisition parameters. We show that deep neural networks trained with an appropriate anatomic context in the network receptive field, can effectively perform ICH segmentation, but those without enough context will overestimate hemorrhage along the skull and around calcifications in the ventricular system. We trained with all data from a multi-center phase II study (n = 112) achieving a best mean and median Dice coefficient of 0.914 and 0.919, a volume correlation of 0.979 and an average volume difference of 1.7 ml and root mean squared error of 4.7 ml in 500 out-of-sample scans from the corresponding multi-center phase III study. 3D networks with appropriate anatomic context outperformed both 2D and random forest models. Our results suggest that deep neural network models, when carefully developed can be incorporated into the workflow of an ICH clinical trial series to quickly and accurately segment ICH, estimate total hemorrhage volume and minimize segmentation failures. The model, weights and scripts for deployment are located at https://github.com/msharrock/deepbleed . This is the first publicly available neural network model for segmentation of ICH, the only model evaluated with the presence of both IVH and SDH and the only model validated in the workflow of a series of clinical trials.

State-of-the-Art Traditional to the Machine- and Deep-Learning-Based Skull Stripping Techniques, Models, and Algorithms

Several neuroimaging processing applications consider skull stripping as a crucial pre-processing step. Due to complex anatomical brain structure and intensity variations in brain magnetic resonance imaging (MRI), an appropriate skull stripping is an important part. The process of skull stripping basically deals with the removal of the skull region for clinical analysis in brain segmentation tasks, and its accuracy and efficiency are quite crucial for diagnostic purposes. It requires more accurate and detailed methods for differentiating brain regions and the skull regions and is considered as a challenging task. This paper is focused on the transition of the conventional to the machine- and deep-learning-based automated skull stripping methods for brain MRI images. It is observed in this study that deep learning approaches have outperformed conventional and machine learning techniques in many ways, but they have their limitations. It also includes the comparative analysis of the current state-of-the-art skull stripping methods, a critical discussion of some challenges, model of quantifying parameters, and future work directions.

Impact of brain volume and intracranial cerebrospinal fluid volume on the clinical outcome in endovascularly treated stroke patients

PURPOSE

Previously, brain volume (BV) and intracranial cerebrospinal fluid volume (CSFV) have been investigated regarding clinical outcomes of subgroups of ischemic stroke patients. This study aimed to examine if the preexisting, preischemic BV and CSFV have an impact on good functional outcome and mortality in acute ischemic stroke (AIS) patients treated with mechanical thrombectomy (MT).

METHODS

Preischemic BV, CSFV, and CSFV/Total intracranial volume (TICV)-ratio were calculated with a fully automated segmentation platform. Univariate and multivariate analyses were used to study associations.

RESULTS

In this retrospective study 107 subsequent AIS patients of a prospective database were included. The segmentation results of the fully automated algorithm based on non-contrast computerized tomography scans (NCCT) correlated significantly with the segmentation results obtained from 3D T1 weighted magnetic resonance images (P < 0.001). In the univariate analysis a preexisting BV (P < 0.001), preexisting CSFV (P = 0.009), and the ratio CSFV/total intracranial volume (P < 0.001) each significantly correlated with good functional outcome and mortality. However, in the multivariate regression analysis, also correcting for patient age, none of these volumes remained to correlate with these outcome parameters.

CONCLUSION

In summary, an association of BV, CSFV, and the CSFV/TICV-ratio with good functional outcome and mortality in AIS treated with MT could not be established. A fully automated segmentation algorithm based on NCCT was successfully developed in-house for calculating the volumes of interest.

Extracting 2D weak labels from volume labels using multiple instance learning in CT hemorrhage detection

Multiple instance learning (MIL) is a supervised learning methodology that aims to allow models to learn instance class labels from bag class labels, where a bag is defined to contain multiple instances. MIL is gaining traction for learning from weak labels but has not been widely applied to 3D medical imaging. MIL is well-suited to clinical CT acquisitions since (1) the highly anisotropic voxels hinder application of traditional 3D networks and (2) patch-based networks have limited ability to learn whole volume labels. In this work, we apply MIL with a deep convolutional neural network to identify whether clinical CT head image volumes possess one or more large hemorrhages (> 20cm3), resulting in a learned 2D model without the need for 2D slice annotations. Individual image volumes are considered separate bags, and the slices in each volume are instances. Such a framework sets the stage for incorporating information obtained in clinical reports to help train a 2D segmentation approach. Within this context, we evaluate the data requirements to enable generalization of MIL by varying the amount of training data. Our results show that a training size of at least 400 patient image volumes was needed to achieve accurate per-slice hemorrhage detection. Over a five-fold cross-validation, the leading model, which made use of the maximum number of training volumes, had an average true positive rate of 98.10%, an average true negative rate of 99.36%, and an average precision of 0.9698. The models have been made available along with source code1 to enabled continued exploration and adaption of MIL in CT neuroimaging.

High-resolution T2-FLAIR and non-contrast CT brain atlas of the elderly

Normative brain atlases are a standard tool for neuroscience research and are, for example, used for spatial normalization of image datasets prior to voxel-based analyses of brain morphology and function. Although many different atlases are publicly available, they are usually biased with respect to an imaging modality and the age distribution. Both effects are well known to negatively impact the accuracy and reliability of the spatial normalization process using non-linear image registration methods. An important and very active neuroscience area that lacks appropriate atlases is lesion-related research in elderly populations (e.g. stroke, multiple sclerosis) for which FLAIR MRI and non-contrast CT are often the clinical imaging modalities of choice. To overcome the lack of atlases for these tasks and modalities, this paper presents high-resolution, age-specific FLAIR and non-contrast CT atlases of the elderly generated using clinical images.

A Method to Estimate Brain Volume from Head CT Images and Application to Detect Brain Atrophy in Alzheimer Disease

BACKGROUND AND PURPOSE

Total brain volume and total intracranial volume are important measures for assessing whole-brain atrophy in Alzheimer disease, dementia, and other neurodegenerative diseases. Unlike MR imaging, which has a number of well-validated fully-automated methods, only a handful of methods segment CT images. Available methods either use enhanced CT, do not estimate both volumes, or require formal validation. Reliable computation of total brain volume and total intracranial volume from CT is needed because head CTs are more widely used than head MRIs in the clinical setting. We present an automated head CT segmentation method (CTseg) to estimate total brain volume and total intracranial volume.

MATERIALS AND METHODS

CTseg adapts a widely used brain MR imaging segmentation method from the Statistical Parametric Mapping toolbox using a CT-based template for initial registration. CTseg was tested and validated using head CT images from a clinical archive.

RESULTS

CTseg showed excellent agreement with 20 manually segmented head CTs. The intraclass correlation was 0.97 (P < .001) for total intracranial volume and 0.94 (P < .001) for total brain volume. When CTseg was applied to a cross-sectional Alzheimer disease dataset (58 with Alzheimer disease patients and 58 matched controls), CTseg detected a loss in percentage total brain volume (as a percentage of total intracranial volume) with age (P < .001) as well as a group difference between patients with Alzheimer disease and controls (P < .01). We observed similar results when total brain volume was modeled with total intracranial volume as a confounding variable.

CONCLUSIONS

In current clinical practice, brain atrophy is assessed by inaccurate and subjective "eyeballing" of CT images. Manual segmentation of head CT images is prohibitively arduous and time-consuming. CTseg can potentially help clinicians to automatically measure total brain volume and detect and track atrophy in neurodegenerative diseases. In addition, CTseg can be applied to large clinical archives for a variety of research studies.

Analysis of different post mortem assessment methods for cerebral edema

While cerebral edema is a live-threatening condition in living persons, also an edema-like fluid redistribution can occur post mortem. In deceased, usually macroscopic signs are evaluated during autopsy in order to determine the presence or absence of cerebral edema. As a quantitative and objective classification is beneficial, an already existing method (Radojevic et al., 2017), which is based on a mathematical formula using the intracranial dimensions and the cerebral weight, was compared to the evaluation of macroscopic signs in 31 cases. The results showed an excellent agreement for the comparison between the raters as well as the measurement methods (at opened skull or in CT images). However, both measurement methods only poorly agree with the macroscopic edema evaluation. In order to find a more concordant method, the normalized cerebral weight, which puts the cerebral weight in relation to the intracranial volume, was calculated for 115 cases. This method resulted in an excellent agreement with the macroscopic rating and showed a clear numerical difference between the edematous and nonedematous group. While the influence of the post mortem time and the cooling time was found to be negligible, the age at death might confound the edema classification due to pre-existing cerebral atrophy leading to lower cerebral weights. In summary, the present study compared different assessment methods to classify cerebral edema and developed a rater independent, objective and quantitative classification method, which was as reliable as the rating of the forensic pathologists.

A Publicly Available, High Resolution, Unbiased CT Brain Template

Clinical imaging relies heavily on X-ray computed tomography (CT) scans for diagnosis and prognosis. Many research applications aim to perform population-level analyses, which require images to be put in the same space, usually defined by a population average, also known as a template. We present an open-source, publicly available, high-resolution CT template. With this template, we provide voxel-wise standard deviation and median images, a basic segmentation of the cerebrospinal fluid spaces, including the ventricles, and a coarse whole brain labeling. This template can be used for spatial normalization of CT scans and research applications, including deep learning. The template was created using an anatomically-unbiased template creation procedure, but is still limited by the population it was derived from, an open CT data set without demographic information. The template and derived images are available at https://github.com/muschellij2/high_res_ct_template.

Distributed deep learning across multisite datasets for generalized CT hemorrhage segmentation

PURPOSE

As deep neural networks achieve more success in the wide field of computer vision, greater emphasis is being placed on the generalizations of these models for production deployment. With sufficiently large training datasets, models can typically avoid overfitting their data; however, for medical imaging it is often difficult to obtain enough data from a single site. Sharing data between institutions is also frequently nonviable or prohibited due to security measures and research compliance constraints, enforced to guard protected health information (PHI) and patient anonymity.

METHODS

In this paper, we implement cyclic weight transfer with independent datasets from multiple geographically disparate sites without compromising PHI. We compare results between single-site learning (SSL) and multisite learning (MSL) models on testing data drawn from each of the training sites as well as two other institutions.

RESULTS

The MSL model attains an average dice similarity coefficient (DSC) of 0.690 on the holdout institution datasets with a volume correlation of 0.914, respectively corresponding to a 7% and 5% statistically significant improvement over the average of both SSL models, which attained an average DSC of 0.646 and average correlation of 0.871.

CONCLUSIONS

We show that a neural network can be efficiently trained on data from two physically remote sites without consolidating patient data to a single location. The resulting network improves model generalization and achieves higher average DSCs on external datasets than neural networks trained on data from a single source.

Physics-informed Deep Learning for Dual-Energy Computed Tomography Image Processing

Dual-energy CT (DECT) was introduced to address the inability of conventional single-energy computed tomography (SECT) to distinguish materials with similar absorbances but different elemental compositions. However, material decomposition algorithms based purely on the physics of the underlying attenuation process have several limitations, leading to low signal-to-noise ratio (SNR) in the derived material-specific images. To overcome these, we trained a convolutional neural network (CNN) to develop a framework to reconstruct non-contrast SECT images from DECT scans. We show that the traditional physics-based decomposition algorithms do not bring to bear the full information content of the image data. A CNN that leverages the underlying physics of the DECT image generation process as well as the anatomic information gleaned via training with actual images can generate higher fidelity processed DECT images.

Machine Learning-Enabled Automated Determination of Acute Ischemic Core From Computed Tomography Angiography

Background and Purpose- The availability of and expertise to interpret advanced neuroimaging recommended in the guideline-based endovascular stroke therapy (EST) evaluation are limited. Here, we develop and validate an automated machine learning-based method that evaluates for large vessel occlusion (LVO) and ischemic core volume in patients using a widely available modality, computed tomography angiogram (CTA). Methods- From our prospectively maintained stroke registry and electronic medical record, we identified patients with acute ischemic stroke and stroke mimics with contemporaneous CTA and computed tomography perfusion (CTP) with RAPID (IschemaView) post-processing as a part of the emergent stroke workup. A novel convolutional neural network named DeepSymNet was created and trained to identify LVO as well as infarct core from CTA source images, against CTP-RAPID definitions. Model performance was measured using 10-fold cross validation and receiver-operative curve area under the curve (AUC) statistics. Results- Among the 297 included patients, 224 (75%) had acute ischemic stroke of which 179 (60%) had LVO. Mean CTP-RAPID ischemic core volume was 23±42 mL. LVO locations included internal carotid artery (13%), M1 (44%), and M2 (21%). The DeepSymNet algorithm autonomously learned to identify the intracerebral vasculature on CTA and detected LVO with AUC 0.88. The method was also able to determine infarct core as defined by CTP-RAPID from the CTA source images with AUC 0.88 and 0.90 (ischemic core ≤30 mL and ≤50 mL). These findings were maintained in patients presenting in early (0-6 hours) and late (6-24 hours) time windows (AUCs 0.90 and 0.91, ischemic core ≤50 mL). DeepSymNet probabilities from CTA images corresponded with CTP-RAPID ischemic core volumes as a continuous variable with r=0.7 (Pearson correlation, P<0.001). Conclusions- These results demonstrate that the information needed to perform the neuroimaging evaluation for endovascular therapy with comparable accuracy to advanced imaging modalities may be present in CTA, and the ability of machine learning to automate the analysis.

Recommendations for Processing Head CT Data

Many research applications of neuroimaging use magnetic resonance imaging (MRI). As such, recommendations for image analysis and standardized imaging pipelines exist. Clinical imaging, however, relies heavily on X-ray computed tomography (CT) scans for diagnosis and prognosis. Currently, there is only one image processing pipeline for head CT, which focuses mainly on head CT data with lesions. We present tools and a complete pipeline for processing CT data, focusing on open-source solutions, that focus on head CT but are applicable to most CT analyses. We describe going from raw DICOM data to a spatially normalized brain within CT presenting a full example with code. Overall, we recommend anonymizing data with Clinical Trials Processor, converting DICOM data to NIfTI using dcm2niix, using BET for brain extraction, and registration using a publicly-available CT template for analysis.

Automated Estimation of Acute Infarct Volume from Noncontrast Head CT Using Image Intensity Inhomogeneity Correction

Identification of early ischemic changes (EIC) on noncontrast head CT scans performed within the first few hours of stroke onset may have important implications for subsequent treatment, though early stroke is poorly delimited on these studies. Lack of sharp lesion boundary delineation in early infarcts precludes manual volume measures, as well as measures using edge-detection or region-filling algorithms. We wished to test a hypothesis that image intensity inhomogeneity correction may provide a sensitive method for identifying the subtle regional hypodensity which is characteristic of early ischemic infarcts. A digital image analysis algorithm was developed using image intensity inhomogeneity correction (IIC) and intensity thresholding. Two different IIC algorithms (FSL and ITK) were compared. The method was evaluated using simulated infarcts and clinical cases. For synthetic infarcts, measured infarct volumes demonstrated strong correlation to the true lesion volume (for 20% decreased density “infarcts,” Pearson r = 0.998 for both algorithms); both algorithms demonstrated improved accuracy with increasing lesion size and decreasing lesion density. In clinical cases (41 acute infarcts in 30 patients), calculated infarct volumes using FSL IIC correlated with the ASPECTS scores (Pearson r = 0.680) and the admission NIHSS (Pearson r = 0.544). Calculated infarct volumes were highly correlated with the clinical decision to treat with IV-tPA. Image intensity inhomogeneity correction, when applied to noncontrast head CT, provides a tool for image analysis to aid in detection of EIC, as well as to evaluate and guide improvements in scan quality for optimal detection of EIC.

Automated brain extraction from head CT and CTA images using convex optimization with shape propagation

BACKGROUND AND OBJECTIVE

Non-Contrast Computer Tomography (NCCT) and CT angiography (CTA) are the most used and widely acceptable imaging modalities in clinical practice for the diagnosis and treatment of acute ischemic stroke (AIS) patients. Brain extraction of CT/CTA images plays an essential role in stroke imaging research. There is no robust automated brain extraction method in the literature that is well established for both NCCT and CTA images. Thus, a validated and automated brain extraction tool for CT imaging would be of great value for both research and clinical practice.

METHODS

The proposed brain extraction method is based on the contour evolution technique, which extracts brain tissues from acquired NCCT and CTA images in a slice-by-slice fashion. Specifically, the proposed approach makes use of a novel propagation framework, which is initialized by a localized slice with the largest brain section in axial views, followed by a geodesic level-set evolution for automatically extracting the brain section in each slice. In particular, the segmented contour propagated from the previous slice is reused to penalize the defined object function for contour evolution to enforce the shape continuity between any two adjacent contours. We show that the defined contour evolution function can be solved iteratively by globally optimal convex optimization.

RESULTS

The proposed brain extraction approach is quantitatively evaluated using 40 NCCT and CTA images acquired from 20 AIS patients and drawn from 4 different vendors, compared to manual segmentations using Dice and Jaccard coefficient metrics. The quantitative results show that the proposed segmentation algorithm is consistently accurate for both NCCT and CTA images using Dice metric. The proposed method is further validated on 1736 NCCT and CTA images of 1331 AIS patients acquired from three multi-national multi-centric clinical trials. A visual check performed on these data demonstrates a low failure rate of 0.4% for 1331 NCCT images and a zero-failure rate for 405 CTA images.

CONCLUSIONS

Both quantitative and qualitative evaluation suggest that the proposed brain extraction approach for NCCT and CTA images can be used for different clinical imaging settings, thus serving to improve current image analysis in the field of neuroimaging.

Thresholds for identifying pathological intracranial pressure in paediatric traumatic brain injury

Intracranial pressure (ICP) monitoring forms an integral part of the management of severe traumatic brain injury (TBI) in children. The prediction of elevated ICP from imaging is important when deciding on whether to implement invasive ICP monitoring for a patient. However, the radiological markers of pathologically elevated ICP have not been specifically validated in paediatric studies. Here in, we describe an objective, non-invasive, quantitative method of stratifying which patients are likely to require invasive monitoring. A retrospective review of patients admitted to Cambridge University Hospital's Paediatric Intensive Care Unit between January 2009 and December 2016 with a TBI requiring invasive neurosurgical monitoring was performed. Radiological biomarkers of TBI (basal cistern volume, ventricular volume, volume of extra-axial haematomas) from CT scans were measured and correlated with epochs of continuous high frequency variables of pressure monitoring around the time of imaging. 38 patients were identified. Basal cistern volume was found to correlate significantly with opening ICP (r = -0.53, p < 0.001). The optimal threshold of basal cistern volume for predicting high ICP ([Formula: see text]20 mmHg) was a relative volume of 0.0055 (sensitivity 79%, specificity 80%). Ventricular volume and extra-axial haematoma volume did not correlate significantly with opening ICP. Our results show that the features of pathologically elevated ICP in children may differ considerably from those validated in adults. The development of quantitative parameters can help to predict which patients would most benefit from invasive neurosurgical monitoring and we present a novel radiological threshold for this.

Distributed deep learning for robust multi-site segmentation of CT imaging after traumatic brain injury

Machine learning models are becoming commonplace in the domain of medical imaging, and with these methods comes an ever-increasing need for more data. However, to preserve patient anonymity it is frequently impractical or prohibited to transfer protected health information (PHI) between institutions. Additionally, due to the nature of some studies, there may not be a large public dataset available on which to train models. To address this conundrum, we analyze the efficacy of transferring the model itself in lieu of data between different sites. By doing so we accomplish two goals: 1) the model gains access to training on a larger dataset that it could not normally obtain and 2) the model better generalizes, having trained on data from separate locations. In this paper, we implement multi-site learning with disparate datasets from the National Institutes of Health (NIH) and Vanderbilt University Medical Center (VUMC) without compromising PHI. Three neural networks are trained to convergence on a computed tomography (CT) brain hematoma segmentation task: one only with NIH data, one only with VUMC data, and one multi-site model alternating between NIH and VUMC data. Resultant lesion masks with the multi-site model attain an average Dice similarity coefficient of 0.64 and the automatically segmented hematoma volumes correlate to those done manually with a Pearson correlation coefficient of 0.87, corresponding to an 8% and 5% improvement, respectively, over the single-site model counterparts.

A Radiodensity Histogram Study of the Brain in Multiple Sclerosis

Multiple sclerosis (MS) is a progressive neurodegenerative disease, affecting 1 million Americans and 2.5 million people globally. Although the diagnosis is made clinically, imaging plays a major role in diagnosing and monitoring disease progression and treatment response. Magnetic resonance imaging (MRI) has proven sensitive in imaging MS lesions, but the characterization offered by routine clinical MRI remains qualitative and with discrepancies between imaging and clinical findings. We investigated the ability of digital analysis of noncontrast head computed tomography (CT) images to detect global brain changes of MS. All routine diagnostic head CTs obtained on patients with known MS obtained from 1 of 2 scan platforms from 6/1/2011 to 6/1/2015 were reviewed. Head CT images from 54 patients with MS met inclusion criteria. Head CT images were processed and histogram metrics were compared to age- and gender- matched control subjects from the same CT scanners during the same time interval. Histogram metrics were correlated with plaque burden as seen on MRI studies. Compared with control subjects, patients had increased total brain radiodensity (P < .0001), further characterized as an increased histogram modal radiodensity (P < .0001) with decrease in histogram skewness (P < .0001). Radiodensity decreased with increasing plaque burden. Similar findings were seen in the patients with only mild plaque burden sub- group. Radiodensity is a unique tissue metric that is not measured by other imaging techniques. Our study finds that brain radiodensity histogram metrics highly correlate with MS, even in cases with minimal plaque burden.

Application of Machine Learning to Automated Analysis of Cerebral Edema in Large Cohorts of Ischemic Stroke Patients

Cerebral edema contributes to neurological deterioration and death after hemispheric stroke but there remains no effective means of preventing or accurately predicting its occurrence. Big data approaches may provide insights into the biologic variability and genetic contributions to severity and time course of cerebral edema. These methods require quantitative analyses of edema severity across large cohorts of stroke patients. We have proposed that changes in cerebrospinal fluid (CSF) volume over time may represent a sensitive and dynamic marker of edema progression that can be measured from routinely available CT scans. To facilitate and scale up such approaches we have created a machine learning algorithm capable of segmenting and measuring CSF volume from serial CT scans of stroke patients. We now present results of our preliminary processing pipeline that was able to efficiently extract CSF volumetrics from an initial cohort of 155 subjects enrolled in a prospective longitudinal stroke study. We demonstrate a high degree of reproducibility in total cranial volume registration between scans (R = 0.982) as well as a strong correlation of baseline CSF volume and patient age (as a surrogate of brain atrophy, R = 0.725). Reduction in CSF volume from baseline to final CT was correlated with infarct volume (R = 0.715) and degree of midline shift (quadratic model, p < 2.2 × 10⁻¹⁶). We utilized generalized estimating equations (GEE) to model CSF volumes over time (using linear and quadratic terms), adjusting for age. This model demonstrated that CSF volume decreases over time (p < 2.2 × 10⁻¹³) and is lower in those with cerebral edema (p = 0.0004). We are now fully automating this pipeline to allow rapid analysis of even larger cohorts of stroke patients from multiple sites using an XNAT (eXtensible Neuroimaging Archive Toolkit) platform. Data on kinetics of edema across thousands of patients will facilitate precision approaches to prediction of malignant edema as well as modeling of variability and further understanding of genetic variants that influence edema severity.