Automated Segmentation Pipeline Research Articles

Multipad agarose plate: a rapid and high-throughput approach for antibiotic susceptibility testing.

We describe a phenotypic antibiotic susceptibility testing (AST) method that can provide an eightfold speed-up in turnaround time compared with the current clinical standard by leveraging advances in microscopy and single-cell imaging. A newly developed growth plate containing 96 agarose pads, termed the multipad agarose plate (MAP), can be assembled at low cost. Pads can be prepared with dilution series of antibiotics. Bacteria are seeded on the pads and automatically imaged using brightfield microscopy, with a fully automated segmentation pipeline quantifying microcolony formation and growth rate. Using a test set of nine antibiotics with very different targets, we demonstrate that accurate minimum inhibitory concentration (MIC) measurements can be performed based on the growth rate of microcolonies within 3 h of incubation with the antibiotic when started from exponential phase. Faster, reliable and high-throughput methods for AST, such as MAP, could improve patient care by expediting treatment initiation and alleviating the burden of antimicrobial resistance.

The Cerebellum and Cognitive Function: Anatomical Evidence from a Transdiagnostic Sample

Multiple lines of evidence across human functional, lesion, and animal data point to a cerebellar role, in particular of crus I, crus II, and lobule VIIB, in cognitive function. However, a mapping of distinct facets of cognitive function to cerebellar structure is missing. We analyzed structural neuroimaging data from the Healthy Brain Network (HBN). Cerebellar parcellation was performed with a validated automated segmentation pipeline (CERES) and stringent visual quality check (n = 662 subjects retained from initial n = 1452). Canonical correlation analyses (CCA) examined regional gray matter volumetric (GMV) differences in association to cognitive function (quantified with NIH Toolbox Cognition domain, NIH-TB), accounting for psychopathology severity, age, sex, scan location, and intracranial volume. Multivariate CCA uncovered a significant correlation between two components entailing a latent cognitive canonical (NIH-TB subscales) and a brain canonical variate (cerebellar GMV and intracranial volume, ICV), surviving bootstrapping and permutation procedures. The components correspond to partly shared cerebellar-cognitive function relationship with a first map encompassing cognitive flexibility (r = 0.89), speed of processing (r = 0.65), and working memory (r = 0.52) associated with regional GMV in crus II (r = 0.57) and lobule X (r = 0.59) and a second map including the crus I (r = 0.49) and lobule VI (r = 0.49) associated with working memory (r = 0.51). We show evidence for a structural subspecialization of the cerebellum topography for cognitive function in a transdiagnostic sample.

Automated segmentation and labeling of subcutaneous mouse implants at 14.1T

Magnetic resonance imaging (MRI) is a valuable tool for studying subcutaneous implants in rodents, providing non-invasive insight into biomaterial conformability and longitudinal characterization. However, considerable variability in existing image analysis techniques, manual segmentation and labeling, as well as the lack of reference atlases as opposed to brain imaging, all render the manual implant segmentation task tedious and extremely time-consuming. To this end, the development of automated and robust segmentation pipelines is a necessary addition to the tools available in rodent imaging research. In this work, we presented and compared commonly used image processing contrast-based segmentation approaches—namely, Canny edge detection, Otsu’s single and multi-threshold methods, and a combination of the latter with morphological operators—with more recently introduced convolutional neural network (CNN-) based models, such as the U-Net and nnU-Net (“no-new-net”). These fully automated end-to-end state-of-the-art neural architectures have shown great promise in online segmentation challenges. We adapted them to the implant segmentation task in mice MRI, with both 2D and 3D implementations. Our results demonstrated the superiority of the 3D nnU-Net model, which is able to robustly segment the implants with an average Dice accuracy of 0.915, and an acceptable absolute volume prediction error of 5.74%. Additionally, we provide researchers in the field with an automated segmentation pipeline in Python, leveraging these CNN-based implementations, and allowing to drastically reduce the manual labeling time from approximately 90 min to less than 5 min (292.959 s ± 6.49 s, N = 30 predictions). The latter addresses the bottleneck of constrained animal experimental time in pre-clinical rodent research.

Understanding the relationship between cerebellar structure and social abilities

BackgroundThe cerebellum contains more than 50% of all neurons in the brain and is involved in a broad range of cognitive functions, including social communication and social cognition. Inconsistent atypicalities in the cerebellum have been reported in individuals with autism compared to controls suggesting the limits of categorical case control comparisons. Alternatively, investigating how clinical dimensions are related to neuroanatomical features, in line with the Research Domain Criteria approach, might be more relevant. We hypothesized that the volume of the “cognitive” lobules of the cerebellum would be associated with social difficulties.MethodsWe analyzed structural MRI data from a large pediatric and transdiagnostic sample (Healthy Brain Network). We performed cerebellar parcellation with a well-validated automated segmentation pipeline (CERES). We studied how social communication abilities—assessed with the social component of the Social Responsiveness Scale (SRS)—were associated with the cerebellar structure, using linear mixed models and canonical correlation analysis.ResultsIn 850 children and teenagers (mean age 10.8 ± 3 years; range 5–18 years), we found a significant association between the cerebellum, IQ and social communication performance in our canonical correlation model.LimitationsCerebellar parcellation relies on anatomical boundaries, which does not overlap with functional anatomy. The SRS was originally designed to identify social impairments associated with autism spectrum disorders.ConclusionOur results unravel a complex relationship between cerebellar structure, social performance and IQ and provide support for the involvement of the cerebellum in social and cognitive processes.

Reduced Cerebellar Volume in Term Infants with Complex Congenital Heart Disease: Correlation with Postnatal Growth Measurements.

: ObjectivesAberrant cerebellar development and the associated neurocognitive deficits has been postulated in infants with congenital heart disease (CHD). Our objective is to investigate the effect of postnatal head and somatic growth on cerebellar development in neonates with CHD. We compared term-born neonates with a history of CHD with a cohort of preterm-born neonates, two cohorts at similar risk for neurodevelopment impairment, in order to determine if they are similarly affected in the early developmental period. Study Design: 51 preterms-born healthy neonates, 62 term-born CHD neonates, and 54 term-born healthy neonates underwent a brain MRI with volumetric imaging. Cerebellar volumes were extracted through an automated segmentation pipeline that was developed in-house. Volumes were correlated with clinical growth parameters at both the birth and time of MRI. Results: The CHD cohort showed significantly lower cerebellar volumes when compared with both the control (p < 0.015) and preterm (p < 0.004) groups. Change in weight from birth to time of MRI showed a moderately strong correlation with cerebellar volume at time of MRI (r = 0.437, p < 0.002) in the preterms, but not in the CHD neonates (r = 0.205, p < 0.116). Changes in birth length and head circumference showed no significant correlation with cerebellar volume at time of MRI in either cohort. Conclusions: Cerebellar development in premature-born infants is associated with change in birth weight in the early post-natal period. This association is not observed in term-born neonates with CHD, suggesting differential mechanisms of aberrant cerebellar development in these perinatal at-risk populations.

Automated 3D Analysis of Clinical Magnetic Resonance Images Demonstrates Significant Reductions in Cam Morphology Following Arthroscopic Intervention in Contrast to Physiotherapy

PurposeTo obtain automated measurements of cam volume, surface area, and height from baseline (preintervention) and 12-month magnetic resonance (MR) images acquired from male and female patients allocated to physiotherapy (PT) or arthroscopic surgery (AS) management for femoroacetabular impingement (FAI) in the Australian FASHIoN trial.MethodsAn automated segmentation pipeline (CamMorph) was used to obtain cam morphology data from three-dimensional (3D) MR hip examinations in FAI patients classified with mild, moderate, or major cam volumes. Pairwise comparisons between baseline and 12-month cam volume, surface area, and height data were performed within the PT and AS patient groups using paired t-tests or Wilcoxon signed-rank tests.ResultsA total of 43 patients were included with 15 PT patients (9 males, 6 females) and 28 AS patients (18 males, 10 females) for premanagement and postmanagement cam morphology assessments. Within the PT male and female patient groups, there were no significant differences between baseline and 12-month mean cam volume (male: 1269 vs 1288 mm3, t[16] = −0.39; female: 545 vs 550 mm,3t[10] = −0.78), surface area (male: 1525 vs 1491 mm2, t[16] = 0.92; female: 885 vs 925 mm,2t[10] = −0.78), maximum height (male: 4.36 vs 4.32 mm, t[16] = 0.34; female: 3.05 vs 2.96 mm, t[10] = 1.05) and average height (male: 2.18 vs 2.18 mm, t[16] = 0.22; female: 1.4 vs 1.43 mm, t[10] = −0.38). In contrast, within the AS male and female patient groups, there were significant differences between baseline and 12-month cam volume (male: 1343 vs 718 mm3, W = 0.0; female: 499 vs 240 mm3, t[18] = 2.89), surface area (male: 1520 vs 1031 mm2, t(34) = 6.48; female: 782 vs 483 mm2, t(18) = 3.02), maximum-height (male: 4.3 vs 3.42 mm, W = 13.5; female: 2.85 vs 2.24 mm, t(18) = 3.04) and average height (male: 2.17 vs 1.52 mm, W = 3.0; female: 1.4 vs 0.94 mm, W = 3.0). In AS patients, 3D bone models provided good visualization of cam bone mass removal postostectomy.ConclusionsAutomated measurement of cam morphology from baseline (preintervention) and 12-month MR images demonstrated that the cam volume, surface area, maximum-height, and average height were significantly smaller in AS patients following ostectomy, whereas there were no significant differences in these cam measures in PT patients from the Australian FASHIoN study.Level of EvidenceLevel II, cohort study.

Towards fully automated segmentation of rat cardiac MRI by leveraging deep learning frameworks

Automated segmentation of human cardiac magnetic resonance datasets has been steadily improving during recent years. Similar applications would be highly useful to improve and speed up the studies of cardiac function in rodents in the preclinical context. However, the transfer of such segmentation methods to the preclinical research is compounded by the limited number of datasets and lower image resolution. In this paper we present a successful application of deep architectures 3D cardiac segmentation for rats in preclinical contexts which to our knowledge has not yet been reported. We developed segmentation models that expand on the standard U-Net architecture and evaluated models separately trained for systole and diastole phases (2MSA) and a single model trained for all phases (1MSA). Furthermore, we calibrated model outputs using a Gaussian process (GP)-based prior to improve phase selection. The resulting models approach human performance in terms of left ventricular segmentation quality and ejection fraction (EF) estimation in both 1MSA and 2MSA settings (Sørensen-Dice score 0.91 ± 0.072 and 0.93 ± 0.032, respectively). 2MSA achieved a mean absolute difference between estimated and reference EF of 3.5 ± 2.5%, while 1MSA resulted in 4.1 ± 3.0%. Applying GPs to 1MSA enabled automating systole and diastole phase selection. Both segmentation approaches (1MSA and 2MSA) were statistically equivalent. Combined with a proposed cardiac phase selection strategy, our work presents an important first step towards a fully automated segmentation pipeline in the context of rat cardiac analysis.

BrainSec: Automated Brain Tissue Segmentation Pipeline for Scalable Neuropathological Analysis.

As neurodegenerative disease pathological hallmarks have been reported in both grey matter (GM) and white matter (WM) with different density distributions, automating the segmentation process of GM/WM would be extremely advantageous for aiding in neuropathologic deep phenotyping. Standard segmentation methods typically involve manual annotations, where a trained researcher traces the delineation of GM/WM in ultra-high-resolution Whole Slide Images (WSIs). This method can be time-consuming and subjective, preventing a scalable analysis on pathology images. This paper proposes an automated segmentation pipeline (BrainSec) combining a Convolutional Neural Network (CNN) module for segmenting GM/WM regions and a post-processing module to remove artifacts/residues of tissues. The final output generates XML annotations that can be visualized via Aperio ImageScope. First, we investigate two baseline models for medical image segmentation: FCN, and U-Net. Then we propose a patch-based approach, BrainSec, to classify the GM/WM/background regions. We demonstrate BrainSec is robust and has reliable performance by testing it on over 180 WSIs that incorporate numerous unique cases as well as distinct neuroanatomic brain regions. We also apply gradient-weighted class activation mapping (Grad-CAM) to interpret the segmentation masks and provide relevant explanations and insights. In addition, we have integrated BrainSec with an existing Amyloid-β pathology classification model into a unified framework (without incurring significant computation complexity) to identify pathologies, visualize their distributions, and quantify each type of pathologies in segmented GM/WM regions, respectively.

Automated segmentation of the locus coeruleus in aging and Alzheimer's disease using 3T neuromelanin-sensitive MRI.

The Locus Coeruleus (LC) may play an important role in the pathogenesis of Alzheimer's disease (AD). Neuromelanin-sensitive Magnetic Resonance Imaging (MRI) permits the visualization of the LC in vivo. Previously, we proposed an automatic, deep learning-based approach to segmenting the LC in a cohort of 82 healthy young and older adults (82 HC; Dünnwald et al. 2021). Here, we applied the same automatic segmentation method to baseline data from the DZNE Longitudinal Cognitive Impairment and Dementia (DELCODE) study comprising 42 healthy elderly adults (HEA), 18 participants with subjective cognitive decline (SCD), 25 participants with mild cognitive impairment (MCI) and 11 participants with AD. The LC was visualized using 3T T1-weighted Fast Low Angle Shot (FLASH) MRI acquired at 0.75mm isotropic resolution. Prior to processing, the data was sinc-interpolated to a resolution of 0.375mm³. The automated segmentation pipeline comprises two consecutive Convolutional Neural Networks (CNNs), which perform LC centerpoint regression and segmentation (see Fig. 1). The networks were trained only on the 82 HC dataset using a cross-validation scheme. We applied the networks to the unseen DELCODE data and assessed their performance with respect to manually delineated masks from two trained raters using Dice similarity coefficient (DSC) and intraclass correlation coefficient (ICC) of contrast ratios (CRs). In HEA, the networks showed similar agreement (DSC, ICC) compared to that observed between two independent raters. Using DSC, automated segmentation of SCD, MCI and AD groups showed less agreement than in the HEA group with 66.67%, 64.86%, 59.77% and 67.06% on average respectively (see Fig. 2). However, good reliability between automated and manually delineated masks for all groups using LC MRI contrast was observed (see Tab. 1) indicating that CRs may be relatively robust wrt. mask differences (see Fig. 3 and 4 for qualitative comparisons). The results demonstrate good reliability between manual and automatically generated LC segmentations with respect to LC MRI contrast using networks trained on healthy older adults. Additional fine-tuning of the networks to account for the variance in AD should be explored as a means to further improve reliability.

Automated segmentation of knee articular cartilage: Joint deep and hand-crafted learning-based framework using diffeomorphic mapping

Segmentation of knee articular cartilage tissue (ACT) from 3D magnetic resonance images (MRIs) is a fundamental task in assessing knee osteoarthritis (KOA). However, automated ACT segmentation of knee cartilage is complicated by (1) the variability of pathological structures in terms of shape, size, spatial resolution and (2) the uncertainties in delineating inter- and intra-cartilage boundaries (possibly touching inter-cartilage surfaces, intensity inhomogeneity, cartilage loss due to disease progression). To address these problems, we propose a novel joint deep and hand-crafted learning-based (JD-HCl) framework. The fully automated segmentation pipeline is a combination of three main steps: 3D convolutional neural networks (CNNs), diffeomorphic mapping, and a conventional hand-crafted feature-based classification model. Firstly, the initial segmentation of knee ACT was drawn from preprocessed knee MRIs using 3D U-Net. Secondly, the spatial alignment in the image domain was carried out using diffeomorphic image registration derived from an estimate of the anatomical correspondences between the subjects and population-specific template. Finally, a tissue-specific hand-crafted feature set was extracted from the spatially aligned pre-segmented ACT region and combined with the semantic context priors obtained from 3D CNN model to develop the second-stage learning model. The proposed method was evaluated with two publicly available Osteoarthritis Initiative (OAI) datasets and one in-house MRI sequences from the Alfred hospital, Melbourne, Australia. The proposed JD-HCl framework produced strong Dice similarity coefficient (DSC), ranging between 97.4% and 98.5% (95% confidence interval) for all cartilage compartments at baseline to 97.2% and 98.5% (the 1-year follow-up) for OAI(-Imorphics) dataset, and 89.3% and 94.6% for OAI-ZIB validation dataset, respectively. Average correlations of cartilage volumes between manual and automatic segmentations were 0.9 and 0.83 for OAI-ZIB dataset, respectively, and 0.975, 0.995, and 0.99 for OAI(-Imorphics) validation datasets, respectively. In longitudinal studies, accurate segmentation of knee ACT using the proposed method produces reproducible cartilage volume, thickness measurements valuable for the study of KOA progression.

Impact of Prenatal Stress on Amygdala Anatomy in Young Adulthood: Timing and Location Matter

BackgroundExposure to maternal stress in utero has long-term implications for the developing brain and has been linked with a higher risk of depression. The amygdala, which develops during the early embryonic stage and is critical for emotion processing, might be particularly sensitive. MethodsUsing data from a neuroimaging follow-up of the European Longitudinal Study of Pregnancy and Childhood prenatal birth cohort (n = 129, 47% men, 23–24 years old), we studied the impact of prenatal stress during the first and second halves of pregnancy on the volume of the amygdala and its nuclei in young adult offspring. We further evaluated the relationship between amygdala anatomy and offspring depressive symptomatology. Amygdala nuclei were parcellated using FreeSurfer’s automated segmentation pipeline. Depressive symptoms were measured via self-report using the Beck Depression Inventory. ResultsExposure to stress during the first half of pregnancy was associated with smaller accessory basal (Cohen’s f2 = 0.27, false discovery rate [FDR]-corrected p [pFDR] = .03) and cortical (Cohen’s f2 = 0.29, pFDR = .03) nuclei volumes. This effect remained significant after correcting for sex, stress during the second half of pregnancy, maternal age at birth, birth weight, maternal education, and offspring’s age at magnetic resonance imaging. These two nuclei showed a quadratic relationship with Beck Depression Inventory scores in young adulthood, where both smaller and larger volumes were associated with more depressive symptoms (accessory basal nucleus: adj. R2 = 0.05, pFDR = .015; cortical nucleus: adj. R2 = 0.04, pFDR = .015). ConclusionsWe conclude that exposure to stress during the first half of pregnancy might have long-term implications for amygdala anatomy, which may in turn predict the experience of depressive symptoms in young adulthood.

The Role of Decreased Cortical Thickness and Volume of Medial Temporal Lobe Structures in Predicting Incident Psychosis in Patients With Alzheimer's Disease: A Prospective Longitudinal MRI Study

To investigate the effect of decreased cortical thickness or volume of medial temporal lobe structures on the risk of incident psychosis in patients with AD. This hospital-based prospective longitudinal study enrolled 109 patients with AD. All patients with AD were evaluated at 3-month intervals to investigate the effect of decreased cortical thickness or volume of medial temporal lobe structures on the risk of incident psychosis in patients with AD. The main outcome measure was time-to-progression from AD to incident psychosis. The thickness or volume of medial temporal lobe structures (i.e., the hippocampus, entorhinal cortex, and parahippocampus) were measured using magnetic resonance imaging and the Freesurfer automated segmentation pipeline at baseline. Multivariate Cox proportional hazards regression analysis revealed that a decreased cortical thickness or volume of medial temporal region was associated with a higher risk of incident psychosis in patients with AD. The hazard ratios for decreased cortical thickness of the left entorhinal cortex and decreased cortical volume of the right hippocampus were 4.291 (95% confidence interval [CI], 1.196-15.384) and 2.680 [(CI, 1.003-1.196]), respectively. Our study revealed that decreased cortical thickness or volume of medial temporal sub-regions is a risk factor for incident psychosis in patients with AD. A careful assessment of the thickness or volume of the medial temporal lobe structures in AD may improve early detection and intervention of psychosis in AD.

Distinct Hippocampal Subfields Atrophy in Older People With Vascular Brain Injuries.

Many neurological or psychiatric diseases affect the hippocampus during aging. The study of hippocampal regional vulnerability may provide important insights into the pathophysiological mechanisms underlying these processes; however, little is known about the specific impact of vascular brain damage on hippocampal subfields atrophy. To analyze the effect of vascular injuries independently of other pathological conditions, we studied a population-based cohort of nondemented older adults, after the exclusion of people who were diagnosed with neurodegenerative diseases during the 14-year clinical follow-up period. Using an automated segmentation pipeline, 1.5T-magnetic resonance imaging at inclusion and 4 years later were assessed to measure both white matter hyperintensities and hippocampal subfields volume. Annualized rates of white matter hyperintensity progression and annualized rates of hippocampal subfields atrophy were then estimated in each participant. We included 249 participants in our analyses (58% women, mean age 71.8, median Mini-Mental State Evaluation 29). The volume of the subiculum at baseline was the only hippocampal subfield volume associated with total, deep/subcortical, and periventricular white matter hyperintensity volumes, independently of demographic variables and vascular risk factors (β=-0.17, P=0.011; β=-0.25, P=0.020 and β=-0.14, P=0.029, respectively). In longitudinal measures, the annualized rate of subiculum atrophy was significantly higher in people with the highest rate of deep/subcortical white matter hyperintensity progression, independently of confounding factors (β=-0.32, P=0.014). These cross-sectional and longitudinal findings highlight the links between vascular brain injuries and a differential vulnerability of the subiculum within the hippocampal loop, unbiased of the effect of neurodegenerative diseases, and particularly when vascular injuries affect deep/subcortical structures.

Fully automated preoperative segmentation of temporal bone structures from clinical CT scans

Middle- and inner-ear surgery is a vital treatment option in hearing loss, infections, and tumors of the lateral skull base. Segmentation of otologic structures from computed tomography (CT) has many potential applications for improving surgical planning but can be an arduous and time-consuming task. We propose an end-to-end solution for the automated segmentation of temporal bone CT using convolutional neural networks (CNN). Using 150 manually segmented CT scans, a comparison of 3 CNN models (AH-Net, U-Net, ResNet) was conducted to compare Dice coefficient, Hausdorff distance, and speed of segmentation of the inner ear, ossicles, facial nerve and sigmoid sinus. Using AH-Net, the Dice coefficient was 0.91 for the inner ear; 0.85 for the ossicles; 0.75 for the facial nerve; and 0.86 for the sigmoid sinus. The average Hausdorff distance was 0.25, 0.21, 0.24 and 0.45 mm, respectively. Blinded experts assessed the accuracy of both techniques, and there was no statistical difference between the ratings for the two methods (p = 0.93). Objective and subjective assessment confirm good correlation between automated segmentation of otologic structures and manual segmentation performed by a specialist. This end-to-end automated segmentation pipeline can help to advance the systematic application of augmented reality, simulation, and automation in otologic procedures.

Robust chest CT image segmentation of COVID-19 lung infection based on limited data.

BackgroundThe coronavirus disease 2019 (COVID-19) affects billions of lives around the world and has a significant impact on public healthcare. For quantitative assessment and disease monitoring medical imaging like computed tomography offers great potential as alternative to RT-PCR methods. For this reason, automated image segmentation is highly desired as clinical decision support. However, publicly available COVID-19 imaging data is limited which leads to overfitting of traditional approaches.MethodsTo address this problem, we propose an innovative automated segmentation pipeline for COVID-19 infected regions, which is able to handle small datasets by utilization as variant databases. Our method focuses on on-the-fly generation of unique and random image patches for training by performing several preprocessing methods and exploiting extensive data augmentation. For further reduction of the overfitting risk, we implemented a standard 3D U-Net architecture instead of new or computational complex neural network architectures.ResultsThrough a k-fold cross-validation on 20 CT scans as training and validation of COVID-19, we were able to develop a highly accurate as well as robust segmentation model for lungs and COVID-19 infected regions without overfitting on limited data. We performed an in-detail analysis and discussion on the robustness of our pipeline through a sensitivity analysis based on the cross-validation and impact on model generalizability of applied preprocessing techniques. Our method achieved Dice similarity coefficients for COVID-19 infection between predicted and annotated segmentation from radiologists of 0.804 on validation and 0.661 on a separate testing set consisting of 100 patients.ConclusionsWe demonstrated that the proposed method outperforms related approaches, advances the state-of-the-art for COVID-19 segmentation and improves robust medical image analysis based on limited data.

A Deep Learning Pipeline to Automate High-Resolution Arterial Segmentation With or Without Intravenous Contrast.

Existing methods to reconstruct vascular structures from a computerized tomography (CT) angiogram rely on contrast injection to enhance the radio-density within the vessel lumen. However, pathological changes in the vasculature may be present that prevent accurate reconstruction. In aortic aneurysmal disease, a thrombus adherent to the aortic wall within the expanding aneurysmal sac is present in >90% of cases. These deformations prevent the automatic extraction of vital clinical information by existing image reconstruction methods. In this study, a deep learning architecture consisting of a modified U-Net with attention-gating was implemented to establish a high-throughput and automated segmentation pipeline of pathological blood vessels in CT images acquired with or without the use of a contrast agent. Seventy-Five patients with paired noncontrast and contrast-enhanced CT images were randomly selected from an ongoing study (Ethics Ref 13/SC/0250), manually annotated and used for model training and evaluation. Data augmentation was implemented to diversify the training data set in a ratio of 10:1. The performance of our Attention-based U-Net in extracting both the inner (blood flow) lumen and the wall structure of the aortic aneurysm from CT angiograms was compared against a generic 3-D U-Net and displayed superior results. Implementation of this network within the aortic segmentation pipeline for both contrast and noncontrast CT images has allowed for accurate and efficient extraction of the morphological and pathological features of the entire aortic volume. This extraction method can be used to standardize aneurysmal disease management and sets the foundation for complex geometric and morphological analysis. Furthermore, this pipeline can be extended to other vascular pathologies.

A deep learning approach to automate high-resolution blood vessel reconstruction on computerised tomography images with or without the use of contrast agents

Abstract Background Existing methods to reconstruct vascular structures from a computed tomography (CT) angiogram rely on injection of intravenous contrast to enhance the radio-density within the vessel lumen. Pathological changes present within the blood lumen, vessel wall or a combination of both prevent accurate 3D reconstruction. In the example of aortic aneurysmal (AAA) disease, a blood clot or thrombus adherent to the aortic wall within the expanding aneurysmal sac is present in 95% of cases. These deformations prevent the automatic extraction of vital clinically relevant information by current methods. Objectives In this study, we utilised deep learning segmentation methods to establish a high-throughput and automated segmentation pipeline for pathological blood vessels (ex. Aortic Aneurysm) in CT images acquired with or without the use of a contrast agent. Methods Twenty-six patients with paired non-contrast and contrast-enhanced CT images were randomly selected from an ethically-approved ongoing study (Ethics Ref 13/SC/0250), manually annotated and used for model training and evaluation (13/13). Data augmentation methods were implemented to diversify the training data set in a ratio of 10:1. We utilised a 3D U-Net with attention gating for both the aortic region-of-interest (ROI) detection and segmentation tasks. Trained architectures were evaluated using the DICE similarity score. Results Inter- and Intra- observer analysis supports the accuracy of the manual segmentations used for model training (intra-class correlation coefficient, “ICC” = 0.995 and 1.00, respective. P&amp;lt;0.001 for both). The performance of our Attention-based U-Net (DICE score: 94.8±0.5%) in extracting both the inner lumen and the outer wall of the aortic aneurysm from CT angiograms (CTA) was compared against a generic 3-D U-Net (DICE score: 89.5±0.6%) and displayed superior results (p&amp;lt;0.01). Fig 1A depicts the implementation of this network architecture within the aortic segmentation pipeline (automated ROI detection and aortic segmentation). This pipeline has allowed accurate and efficient extraction of the entire aortic volume from both contrast-enhanced CTA (DICE score: 95.3±0.6%) and non-contrast CT (DICE score: 93.2±0.7%) images. Fig 1B illustrates the model output alongside the labelled ground truth segmentation for the pathological aneurysmal region; only minor differences are visually discernible (coloured boxes). Conclusion We developed a novel automated pipeline for high resolution reconstruction of blood vessels using deep learning approaches. This pipeline enables automatic extraction of morphologic features of blood vessels and can be applied for research and potentially for clinical use. Automated Segmentation of Blood Vessels Funding Acknowledgement Type of funding source: Foundation. Main funding source(s): University of Oxford Medical Research Fund, John Fell Fund

Bone-Selective Magnetic Resonance Imaging as a Nonradiative Alternative to Computed Tomography for Cranial Vault Imaging: Concordance and Implementation of an Automated Segmentation Pipeline for Timely Image Processing

BACKGROUND: Computed tomography (CT) is the clinical gold standard for high-resolution 3-dimensional (3D) visualization of cortical bone structures. However, CT ionizing radiation exposure is associated with the development of malignancy. Bone-selective magnetic resonance imaging (MRI) and bone-selective image reconstruction provide a radiation-free imaging modality with diagnostic and surgical planning uses. This technique, though applicable to many realms of plastic and reconstructive surgery, is of specific interest to craniofacial surgeons whose patients often require multiple pre- and postoperative CT scans, enduring a higher cumulative risk of malignancy. As it stands, the implementation of bone-selective MRI in clinical practice is prevented by a paucity of CT and bone-selective MRI concordance data and the time and labor intensive process required to produce bone-selective MR-based 3D skull segmentations. The manual segmentation process takes about 1.5 hours of time per MRI. Our study evaluates both the accuracy of a novel bone-selective MRI technique (dual-radiofrequency pulse, dual-echo, 3D ultrashort echo time) and the utility of a segmentation pipeline. OBJECTIVES: Part 1. Evaluate the concordance between MR-based and CT-based 3D skull renderings Part 2. Describe and evaluate a novel multiatlas segmentation pipeline DESIGN/METHODS: Part 1: A cadaver skull and the skulls of 5 healthy adult volunteers were scanned with bone-selective MR and thin-slice CT. Semi-automatic bone segmentation (1.5 hours/scan) was performed creating 3D renderings of the skulls. Mimics software was used to measure 8 anatomic distances from the 3D renderings. Lin’s Concordance Correlation test was applied to assess agreement between MR and CT-based 3D renderings. Part 2. CT and bone-selective MR images were acquired from 16 additional healthy adult volunteers, yielding 21 MR/CT pairs. The CT images were segmented using a semi-automated method to generate “ground truth” labels for the MR images. An automated multiatlas segmentation pipeline was then used to segment the 3D MR images using a 2-step process consisting of a training and segmentation. The training step develops an “atlas package,” which represents the varying anatomy from different subjects. The segmentation step uses the atlas package to generate segmentations for new subjects using several image registration steps. RESULTS: MR-based measurements differed from CT-based measurements by mean percent difference ranging from 2.3% to 5.0%. Lin’s Concordance Correlation ranged from 0.998 to 1.000. The segmentation pipeline took 10 minutes per segmentation with an average symmetric surface distance of 0.96 ± 0.15 mm between the manual reference segmentation and the corresponding automated segmentations. CONCLUSIONS: This study demonstrates high concordance between the gold standard (thin-slice CT) and our novel imaging modality as well as an 89% reduction in segmentation time. This technique is highly applicable to craniofacial surgery as well as cases involving extremity surgery, musculoskeletal trauma, and bone tumors. It additionally allows acquision of data of both soft and hard tissue structures from a single imaging modality with no radiation exposure. The demonstrated reduced segmentation time would allow bone-selective MRI to be used in clinical practice without a delay in treatment. We plan to investigate the accuracy of this technique as a tool for craniosynostosis diagnosis as well as in craniofacial virtual surgical planning.

CT-based True- and False-Lumen Segmentation in Type B Aortic Dissection Using Machine Learning.

To develop a segmentation pipeline for segmentation of aortic dissection CT angiograms into true and false lumina on multiplanar reformations (MPRs) perpendicular to the aortic centerline and derive quantitative morphologic features, specifically aortic diameter and true- or false-lumen cross-sectional area. An automated segmentation pipeline including two convolutional neural network (CNN) segmentation algorithms was developed. The algorithm derives the aortic centerline, generates MPRs orthogonal to the centerline, and segments the true and false lumina. A total of 153 CT angiograms obtained from 45 retrospectively identified patients (mean age, 50 years; range, 22-79 years) were used to train (n = 103), validate (n = 22), and test (n = 28) the CNN pipeline. Accuracy was evaluated by using the Dice similarity coefficient (DSC). Segmentations were then used to derive the maximal diameter of test-set patients and cross-sectional area profiles of the true and false lumina. The segmentation pipeline yielded a mean DSC of 0.873 ± 0.056 for the true lumina and 0.894 ± 0.040 for the false lumina of test-set cases. Automated maximal diameter measurements correlated well with manual measurements (R 2 = 0.95). Profiles of cross-sectional diameter, true-lumen area, and false-lumen area over several follow-up examinations were derived. A segmentation pipeline was used to accurately identify true and false lumina on CT angiograms of aortic dissection. These segmentations can be used to obtain diameter and other morphologic parameters for surveillance and risk stratification.Supplemental material is available for this article.© RSNA, 2020.

An Automated Segmentation Pipeline for Intratumoural Regions in Animal Xenografts Using Machine Learning and Saturation Transfer MRI

Saturation transfer MRI can be useful in the characterization of different tumour types. It is sensitive to tumour metabolism, microstructure, and microenvironment. This study aimed to use saturation transfer to differentiate between intratumoural regions, demarcate tumour boundaries, and reduce data acquisition times by identifying the imaging scheme with the most impact on segmentation accuracy. Saturation transfer-weighted images were acquired over a wide range of saturation amplitudes and frequency offsets along with T1 and T2 maps for 34 tumour xenografts in mice. Independent component analysis and Gaussian mixture modelling were used to segment the images and identify intratumoural regions. Comparison between the segmented regions and histopathology indicated five distinct clusters: three corresponding to intratumoural regions (active tumour, necrosis/apoptosis, and blood/edema) and two extratumoural (muscle and a mix of muscle and connective tissue). The fraction of tumour voxels segmented as necrosis/apoptosis quantitatively matched those calculated from TUNEL histopathological assays. An optimal protocol was identified providing reasonable qualitative agreement between MRI and histopathology and consisting of T1 and T2 maps and 22 magnetization transfer (MT)-weighted images. A three-image subset was identified that resulted in a greater than 90% match in positive and negative predictive value of tumour voxels compared to those found using the entire 24-image dataset. The proposed algorithm can potentially be used to develop a robust intratumoural segmentation method.