Knowing What You Know in Brain Segmentation Using Bayesian Deep Neural Networks

Trustworthy clinical AI solutions: A unified review of uncertainty quantification in Deep Learning models for medical image analysis

The full acceptance of Deep Learning (DL) models in the clinical field is rather low with respect to the quantity of high-performing solutions reported in the literature. End users are particularly reluctant to rely on the opaque predictions of DL models. Uncertainty quantification methods have been proposed in the literature as a potential solution, to reduce the black-box effect of DL models and increase the interpretability and the acceptability of the result by the final user. In this review, we propose an overview of the existing methods to quantify uncertainty associated with DL predictions. We focus on applications to medical image analysis, which present specific challenges due to the high dimensionality of images and their variable quality, as well as constraints associated with real-world clinical routine. Moreover, we discuss the concept of structural uncertainty, a corpus of methods to facilitate the alignment of segmentation uncertainty estimates with clinical attention. We then discuss the evaluation protocols to validate the relevance of uncertainty estimates. Finally, we highlight the open challenges for uncertainty quantification in the medical field.

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.

CHIASM-Net: Artificial Intelligence-Based Direct Identification of Chiasmal Abnormalities in Albinism

Purpose

Albinism is a congenital disorder affecting pigmentation levels, structure, and function of the visual system. The identification of anatomical changes typical for people with albinism (PWA), such as optic chiasm malformations, could become an important component of diagnostics. Here, we tested an application of convolutional neural networks (CNNs) for this purpose.

Methods

We established and evaluated a CNN, referred to as CHIASM-Net, for the detection of chiasmal malformations from anatomic magnetic resonance (MR) images of the brain. CHIASM-Net, composed of encoding and classification modules, was developed using MR images of controls (n = 1708) and PWA (n = 32). Evaluation involved 8-fold cross validation involving accuracy, precision, recall, and F1-score metrics and was performed on a subset of controls and PWA samples excluded from the training. In addition to quantitative metrics, we used Explainable AI (XAI) methods that granted insights into factors driving the predictions of CHIASM-Net.

Results

The results for the scenario indicated an accuracy of 85 ± 14%, precision of 90 ± 14% and recall of 81 ± 18%. XAI methods revealed that the predictions of CHIASM-Net are driven by optic-chiasm white matter and by the optic tracts.

Conclusions

CHIASM-Net was demonstrated to use relevant regions of the optic chiasm for albinism detection from magnetic resonance imaging (MRI) brain anatomies. This indicates the strong potential of CNN-based approaches for visual pathway analysis and ultimately diagnostics.

A network score-based metric to optimize the quality assurance of automatic radiotherapy target segmentations

Background and purpose

Existing methods for quality assurance of the radiotherapy auto-segmentations focus on the correlation between the average model entropy and the Dice Similarity Coefficient (DSC) only. We identified a metric directly derived from the output of the network and correlated it with clinically relevant metrics for contour accuracy.

Materials and Methods

Magnetic Resonance Imaging auto-segmentations were available for the gross tumor volume for cervical cancer brachytherapy (106 segmentations) and for the clinical target volume for rectal cancer external-beam radiotherapy (77 segmentations). The nnU-Net's output before binarization was taken as a score map. We defined a metric as the mean of the voxels in the score map above a threshold (λ). Comparisons were made with the mean and standard deviation over the score map and with the mean over the entropy map. The DSC, the 95th Hausdorff distance, the mean surface distance (MSD) and the surface DSC were computed for segmentation quality. Correlations between the studied metrics and model quality were assessed with the Pearson correlation coefficient (r). The area under the curve (AUC) was determined for detecting segmentations that require reviewing.

Results

For both tasks, our metric (λ = 0.30) correlated more strongly with the segmentation quality than the mean over the entropy map (for surface DSC, r > 0.65 vs. r < 0.60). The AUC was above 0.84 for detecting MSD values above 2 mm.

Conclusions

Our metric correlated strongly with clinically relevant segmentation metrics and detected segmentations that required reviewing, indicating its potential for automatic quality assurance of radiotherapy target auto-segmentations.

Brain Growth Charts for Quantitative Analysis of Pediatric Clinical Brain MRI Scans with Limited Imaging Pathology

Background Clinically acquired brain MRI scans represent a valuable but underused resource for investigating neurodevelopment due to their technical heterogeneity and lack of appropriate controls. These barriers have curtailed retrospective studies of clinical brain MRI scans compared with more costly prospectively acquired research-quality brain MRI scans. Purpose To provide a benchmark for neuroanatomic variability in clinically acquired brain MRI scans with limited imaging pathology (SLIPs) and to evaluate if growth charts from curated clinical MRI scans differed from research-quality MRI scans or were influenced by clinical indication for the scan. Materials and Methods In this secondary analysis of preexisting data, clinical brain MRI SLIPs from an urban pediatric health care system (individuals aged ≤22 years) were scanned across nine 3.0-T MRI scanners. The curation process included manual review of signed radiology reports and automated and manual quality review of images without gross pathology. Global and regional volumetric imaging phenotypes were measured using two image segmentation pipelines, and clinical brain growth charts were quantitatively compared with charts derived from a large set of research controls in the same age range by means of Pearson correlation and age at peak volume. Results The curated clinical data set included 532 patients (277 male; median age, 10 years [IQR, 5-14 years]; age range, 28 days after birth to 22 years) scanned between 2005 and 2020. Clinical brain growth charts were highly correlated with growth charts derived from research data sets (22 studies, 8346 individuals [4947 male]; age range, 152 days after birth to 22 years) in terms of normative developmental trajectories predicted by the models (median r = 0.979). Conclusion The clinical indication of the scans did not significantly bias the output of clinical brain charts. Brain growth charts derived from clinical controls with limited imaging pathology were highly correlated with brain charts from research controls, suggesting the potential of curated clinical MRI scans to supplement research data sets. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Ertl-Wagner and Pai in this issue.

Impact of brain segmentation methods on regional metabolism quantification in 18F-FDG PET/MR analysis

BACKGROUND

Accurate analysis of quantitative PET data plays a crucial role in studying small, specific brain structures. The integration of PET and MRI through an integrated PET/MR system presents an opportunity to leverage the benefits of precisely aligned structural MRI and molecular PET images in both spatial and temporal dimensions. However, in many clinical workflows, PET studies are often performed without the aid of individually matched structural MRI scans, primarily for the sake of convenience in the data collection and brain segmentation possesses. Currently, two commonly employed segmentation strategies for brain PET analysis are distinguished: methods with or without MRI registration and methods employing either atlas-based or individual-based algorithms. Moreover, the development of artificial intelligence (AI)-assisted methods for predicting brain segmentation holds promise but requires further validation of their efficiency and accuracy for clinical applications. This study aims to compare and evaluate the correlations, consistencies, and differences among the above-mentioned brain segmentation strategies in quantification of brain metabolism in 18F-FDG PET/MR analysis.

RESULTS

Strong correlations were observed among all methods (r = 0.932 to 0.999, P < 0.001). The variances attributable to subject and brain region were higher than those caused by segmentation methods (P < 0.001). However, intraclass correlation coefficient (ICC)s between methods with or without MRI registration ranged from 0.924 to 0.975, while ICCs between methods with atlas- or individual-based algorithms ranged from 0.741 to 0.879. Brain regions exhibiting significant standardized uptake values (SUV) differences due to segmentation methods were the basal ganglia nuclei (maximum to 11.50 ± 4.67%), and various cerebral cortexes in temporal and occipital regions (maximum to 18.03 ± 5.52%). The AI-based method demonstrated high correlation (r = 0.998 and 0.999, P < 0.001) and ICC (0.998 and 0.997) with FreeSurfer, substantially reducing the time from 8.13 h to 57 s on per subject.

CONCLUSIONS

Different segmentation methods may have impact on the calculation of brain metabolism in basal ganglia nuclei and specific cerebral cortexes. The AI-based approach offers improved efficiency and is recommended for its enhanced performance.

Comparative validation of AI and non-AI methods in MRI volumetry to diagnose Parkinsonian syndromes

Automated segmentation and volumetry of brain magnetic resonance imaging (MRI) scans are essential for the diagnosis of Parkinson's disease (PD) and Parkinson's plus syndromes (P-plus). To enhance the diagnostic performance, we adopt deep learning (DL) models in brain MRI segmentation and compared their performance with the gold-standard non-DL method. We collected brain MRI scans of healthy controls ([Formula: see text]) and patients with PD ([Formula: see text]), multiple systemic atrophy ([Formula: see text]), and progressive supranuclear palsy ([Formula: see text]) at Samsung Medical Center from January 2017 to December 2020. Using the gold-standard non-DL model, FreeSurfer (FS), we segmented six brain structures: midbrain, pons, caudate, putamen, pallidum, and third ventricle, and considered them as annotated data for DL models, the representative convolutional neural network (CNN) and vision transformer (ViT)-based models. Dice scores and the area under the curve (AUC) for differentiating normal, PD, and P-plus cases were calculated to determine the measure to which FS performance can be reproduced as-is while increasing speed by the DL approaches. The segmentation times of CNN and ViT for the six brain structures per patient were 51.26 ± 2.50 and 1101.82 ± 22.31 s, respectively, being 14 to 300 times faster than FS (15,735 ± 1.07 s). Dice scores of both DL models were sufficiently high (> 0.85) so their AUCs for disease classification were not inferior to that of FS. For classification of normal vs. P-plus and PD vs. P-plus (except multiple systemic atrophy - Parkinsonian type) based on all brain parts, the DL models and FS showed AUCs above 0.8, demonstrating the clinical value of DL models in addition to FS. DL significantly reduces the analysis time without compromising the performance of brain segmentation and differential diagnosis. Our findings may contribute to the adoption of DL brain MRI segmentation in clinical settings and advance brain research.

Application of simultaneous uncertainty quantification for image segmentation with probabilistic deep learning: Performance benchmarking of oropharyngeal cancer target delineation as a use-case

Background

Oropharyngeal cancer (OPC) is a widespread disease, with radiotherapy being a core treatment modality. Manual segmentation of the primary gross tumor volume (GTVp) is currently employed for OPC radiotherapy planning, but is subject to significant interobserver variability. Deep learning (DL) approaches have shown promise in automating GTVp segmentation, but comparative (auto)confidence metrics of these models predictions has not been well-explored. Quantifying instance-specific DL model uncertainty is crucial to improving clinician trust and facilitating broad clinical implementation. Therefore, in this study, probabilistic DL models for GTVp auto-segmentation were developed using large-scale PET/CT datasets, and various uncertainty auto-estimation methods were systematically investigated and benchmarked.

Methods

We utilized the publicly available 2021 HECKTOR Challenge training dataset with 224 co-registered PET/CT scans of OPC patients with corresponding GTVp segmentations as a development set. A separate set of 67 co-registered PET/CT scans of OPC patients with corresponding GTVp segmentations was used for external validation. Two approximate Bayesian deep learning methods, the MC Dropout Ensemble and Deep Ensemble, both with five submodels, were evaluated for GTVp segmentation and uncertainty performance. The segmentation performance was evaluated using the volumetric Dice similarity coefficient (DSC), mean surface distance (MSD), and Hausdorff distance at 95% (95HD). The uncertainty was evaluated using four measures from literature: coefficient of variation (CV), structure expected entropy, structure predictive entropy, and structure mutual information, and additionally with our novel Dice-risk measure. The utility of uncertainty information was evaluated with the accuracy of uncertainty-based segmentation performance prediction using the Accuracy vs Uncertainty (AvU) metric, and by examining the linear correlation between uncertainty estimates and DSC. In addition, batch-based and instance-based referral processes were examined, where the patients with high uncertainty were rejected from the set. In the batch referral process, the area under the referral curve with DSC (R-DSC AUC) was used for evaluation, whereas in the instance referral process, the DSC at various uncertainty thresholds were examined.

Results

Both models behaved similarly in terms of the segmentation performance and uncertainty estimation. Specifically, the MC Dropout Ensemble had 0.776 DSC, 1.703 mm MSD, and 5.385 mm 95HD. The Deep Ensemble had 0.767 DSC, 1.717 mm MSD, and 5.477 mm 95HD. The uncertainty measure with the highest DSC correlation was structure predictive entropy with correlation coefficients of 0.699 and 0.692 for the MC Dropout Ensemble and the Deep Ensemble, respectively. The highest AvU value was 0.866 for both models. The best performing uncertainty measure for both models was the CV which had R-DSC AUC of 0.783 and 0.782 for the MC Dropout Ensemble and Deep Ensemble, respectively. With referring patients based on uncertainty thresholds from 0.85 validation DSC for all uncertainty measures, on average the DSC improved from the full dataset by 4.7% and 5.0% while referring 21.8% and 22% patients for MC Dropout Ensemble and Deep Ensemble, respectively.

Conclusion

We found that many of the investigated methods provide overall similar but distinct utility in terms of predicting segmentation quality and referral performance. These findings are a critical first-step towards more widespread implementation of uncertainty quantification in OPC GTVp segmentation.

Towards a safe and efficient clinical implementation of machine learning in radiation oncology by exploring model interpretability, explainability and data-model dependency

The interest in machine learning (ML) has grown tremendously in recent years, partly due to the performance leap that occurred with new techniques of deep learning, convolutional neural networks for images, increased computational power, and wider availability of large datasets. Most fields of medicine follow that popular trend and, notably, radiation oncology is one of those that are at the forefront, with already a long tradition in using digital images and fully computerized workflows. ML models are driven by data, and in contrast with many statistical or physical models, they can be very large and complex, with countless generic parameters. This inevitably raises two questions, namely, the tight dependence between the models and the datasets that feed them, and the interpretability of the models, which scales with its complexity. Any problems in the data used to train the model will be later reflected in their performance. This, together with the low interpretability of ML models, makes their implementation into the clinical workflow particularly difficult. Building tools for risk assessment and quality assurance of ML models must involve then two main points: interpretability and data-model dependency. After a joint introduction of both radiation oncology and ML, this paper reviews the main risks and current solutions when applying the latter to workflows in the former. Risks associated with data and models, as well as their interaction, are detailed. Next, the core concepts of interpretability, explainability, and data-model dependency are formally defined and illustrated with examples. Afterwards, a broad discussion goes through key applications of ML in workflows of radiation oncology as well as vendors' perspectives for the clinical implementation of ML.

Personalized tDCS for Focal Epilepsy—A Narrative Review: A Data-Driven Workflow Based on Imaging and EEG Data

Conventional transcranial electric stimulation(tES) using standard anatomical positions for the electrodes and standard stimulation currents is frequently not sufficiently selective in targeting and reaching specific brain locations, leading to suboptimal application of electric fields. Recent advancements in in vivo electric field characterization may enable clinical researchers to derive better relationships between the electric field strength and the clinical results. Subject-specific electric field simulations could lead to improved electrode placement and more efficient treatments. Through this narrative review, we present a processing workflow to personalize tES for focal epilepsy, for which there is a clear cortical target to stimulate. The workflow utilizes clinical imaging and electroencephalography data and enables us to relate the simulated fields to clinical outcomes. We review and analyze the relevant literature for the processing steps in the workflow, which are the following: tissue segmentation, source localization, and stimulation optimization. In addition, we identify shortcomings and ongoing trends with regard to, for example, segmentation quality and tissue conductivity measurements. The presented processing steps result in personalized tES based on metrics like focality and field strength, which allow for correlation with clinical outcomes.

FastSurferVINN: Building resolution-independence into deep learning segmentation methods-A solution for HighRes brain MRI

Leading neuroimaging studies have pushed 3T MRI acquisition resolutions below 1.0 mm for improved structure definition and morphometry. Yet, only few, time-intensive automated image analysis pipelines have been validated for high-resolution (HiRes) settings. Efficient deep learning approaches, on the other hand, rarely support more than one fixed resolution (usually 1.0 mm). Furthermore, the lack of a standard submillimeter resolution as well as limited availability of diverse HiRes data with sufficient coverage of scanner, age, diseases, or genetic variance poses additional, unsolved challenges for training HiRes networks. Incorporating resolution-independence into deep learning-based segmentation, i.e., the ability to segment images at their native resolution across a range of different voxel sizes, promises to overcome these challenges, yet no such approach currently exists. We now fill this gap by introducing a Voxel-size Independent Neural Network (VINN) for resolution-independent segmentation tasks and present FastSurferVINN, which (i) establishes and implements resolution-independence for deep learning as the first method simultaneously supporting 0.7-1.0 mm whole brain segmentation, (ii) significantly outperforms state-of-the-art methods across resolutions, and (iii) mitigates the data imbalance problem present in HiRes datasets. Overall, internal resolution-independence mutually benefits both HiRes and 1.0 mm MRI segmentation. With our rigorously validated FastSurferVINN we distribute a rapid tool for morphometric neuroimage analysis. The VINN architecture, furthermore, represents an efficient resolution-independent segmentation method for wider application.

Effect of head motion-induced artefacts on the reliability of deep learning-based whole-brain segmentation

Due to their robustness and speed, recently developed deep learning-based methods have the potential to provide a faster and hence more scalable alternative to more conventional neuroimaging analysis pipelines in terms of whole-brain segmentation based on magnetic resonance (MR) images. These methods were also shown to have higher test-retest reliability, raising the possibility that they could also exhibit superior head motion tolerance. We investigated this by comparing the effect of head motion-induced artifacts in structural MR images on the consistency of segmentation performed by FreeSurfer and recently developed deep learning-based methods to a similar extent. We used state-of-the art neural network models (FastSurferCNN and Kwyk) and developed a new whole-brain segmentation pipeline (ReSeg) to examine whether reliability depends on choice of deep learning method. Structural MRI scans were collected from 110 participants under rest and active head motion and were evaluated for image quality by radiologists. Compared to FreeSurfer, deep learning-based methods provided more consistent segmentations across different levels of image quality, suggesting that they also have the advantage of providing more reliable whole-brain segmentations of MR images corrupted by motion-induced artifacts, and provide evidence for their practical applicability in the study of brain structural alterations in health and disease.

CEREBRUM-7T: Fast and Fully Volumetric Brain Segmentation of 7 Tesla MR Volumes

Ultra-high-field magnetic resonance imaging (MRI) enables sub-millimetre resolution imaging of the human brain, allowing the study of functional circuits of cortical layers at the meso-scale. An essential step in many functional and structural neuroimaging studies is segmentation, the operation of partitioning the MR images in anatomical structures. Despite recent efforts in brain imaging analysis, the literature lacks in accurate and fast methods for segmenting 7-tesla (7T) brain MRI. We here present CEREBRUM-7T, an optimised end-to-end convolutional neural network, which allows fully automatic segmentation of a whole 7T T1_w MRI brain volume at once, without partitioning the volume, pre-processing, nor aligning it to an atlas. The trained model is able to produce accurate multi-structure segmentation masks on six different classes plus background in only a few seconds. The experimental part, a combination of objective numerical evaluations and subjective analysis, confirms that the proposed solution outperforms the training labels it was trained on and is suitable for neuroimaging studies, such as layer functional MRI studies. Taking advantage of a fine-tuning operation on a reduced set of volumes, we also show how it is possible to effectively apply CEREBRUM-7T to different sites data. Furthermore, we release the code, 7T data, and other materials, including the training labels and the Turing test.

Deep Learning-Based Detection of Malformed Optic Chiasms From MRI Images

Convolutional neural network (CNN) models are of great promise to aid the segmentation and analysis of brain structures. Here, we tested whether CNN trained to segment normal optic chiasms from the T1w magnetic resonance imaging (MRI) image can be also applied to abnormal chiasms, specifically with optic nerve misrouting as typical for human albinism. We performed supervised training of the CNN on the T1w images of control participants (n = 1049) from the Human Connectome Project (HCP) repository and automatically generated algorithm-based optic chiasm masks. The trained CNN was subsequently tested on data of persons with albinism (PWA; n = 9) and controls (n = 8) from the CHIASM repository. The quality of outcome segmentation was assessed via the comparison to manually defined optic chiasm masks using the Dice similarity coefficient (DSC). The results revealed contrasting quality of masks obtained for control (mean DSC ± SEM = 0.75 ± 0.03) and PWA data (0.43 ± 0.8, few-corrected p = 0.04). The fact that the CNN recognition of the optic chiasm fails for chiasm abnormalities in PWA underlines the fundamental differences in their spatial features. This finding provides proof of concept for a novel deep-learning-based diagnostics approach of chiasmal misrouting from T1w images, as well as further analyses on chiasmal misrouting and their impact on the structure and function of the visual system.

3D hemisphere-based convolutional neural network for whole-brain MRI segmentation

Whole-brain segmentation is a crucial pre-processing step for many neuroimaging analyses pipelines. Accurate and efficient whole-brain segmentations are important for many neuroimage analysis tasks to provide clinically relevant information. Several recently proposed convolutional neural networks (CNN) perform whole brain segmentation using individual 2D slices or 3D patches as inputs due to graphical processing unit (GPU) memory limitations, and use sliding windows to perform whole brain segmentation during inference. However, these approaches lack global and spatial information about the entire brain and lead to compromised efficiency during both training and testing. We introduce a 3D hemisphere-based CNN for automatic whole-brain segmentation of T1-weighted magnetic resonance images of adult brains. First, we trained a localization network to predict bounding boxes for both hemispheres. Then, we trained a segmentation network to segment one hemisphere, and segment the opposing hemisphere by reflecting it across the mid-sagittal plane. Our network shows high performance both in terms of segmentation efficiency and accuracy (0.84 overall Dice similarity and 6.1 mm overall Hausdorff distance) in segmenting 102 brain structures. On multiple independent test datasets, our method demonstrated a competitive performance in the subcortical segmentation task and a high consistency in volumetric measurements of intra-session scans.

Uncertainty Estimation for Deep Learning-Based Segmentation of Roads in Synthetic Aperture Radar Imagery

Mission-critical applications that rely on deep learning (DL) for automation suffer because DL models struggle to provide reliable indicators of failure. Reliable failure prediction can greatly improve the efficiency of a system, because it becomes easier to predict when human intervention is required. DL-based systems thus stand to benefit greatly from robust measures of uncertainty over model predictions. Monte Carlo dropout (MCD), a Bayesian method, and deep ensembles (DE) have emerged as two of the most popular and competitive ways to perform uncertainty estimation. Although literature exploring the usefulness of these approaches exists in medical imaging, robotics and autonomous driving domains, it is scarce to non-existent for remote sensing, and in particular, synthetic aperture radar (SAR) applications. To close this gap, we have created a deep learning model for road extraction (hereafter referred to as segmentation) in SAR and use it to compare standard model outputs against the aforementioned most popular methods for uncertainty estimation, MCD and DE. We demonstrate that these methods are not effective as an indicator of segmentation quality when measuring uncertainty (as indicated by model softmax outputs) across an entire image but are effective when uncertainty is measured from the set of road predictions only. Furthermore, we show a marked improvement in the correlation between prediction uncertainty and segmentation quality when we increase the set of road predictions by including predictions with lower softmax scores. We demonstrate the efficacy of our application of MCD and DE methods with an experimental design that measures performance in real-world quality assessment using in-distribution (ID) and out-of-distribution (OOD) data. These results inform the development of mission-critical deep learning systems in remote sensing. Tasks in medical image analysis that have a similar morphology to road structures, such as blood vessel segmentation, can also benefit from our findings.

Comparison of Multispectral Image-Processing Methods for Brain Tissue Classification in BrainWeb Synthetic Data and Real MR Images

Accurate quantification of brain tissue is a fundamental and challenging task in neuroimaging. Over the past two decades, statistical parametric mapping (SPM) and FMRIB's Automated Segmentation Tool (FAST) have been widely used to estimate gray matter (GM) and white matter (WM) volumes. However, they cannot reliably estimate cerebrospinal fluid (CSF) volumes. To address this problem, we developed the TRIO algorithm (TRIOA), a new magnetic resonance (MR) multispectral classification method. SPM8, SPM12, FAST, and the TRIOA were evaluated using the BrainWeb database and real magnetic resonance imaging (MRI) data. In this paper, the MR brain images of 140 healthy volunteers (51.5 ± 15.8 y/o) were obtained using a whole-body 1.5 T MRI system (Aera, Siemens, Erlangen, Germany). Before classification, several preprocessing steps were performed, including skull stripping and motion and inhomogeneity correction. After extensive experimentation, the TRIOA was shown to be more effective than SPM and FAST. For real data, all test methods revealed that the participants aged 20–83 years exhibited an age-associated decline in GM and WM volume fractions. However, for CSF volume estimation, SPM8-s and SPM12-m both produced different results, which were also different compared with those obtained by FAST and the TRIOA. Furthermore, the TRIOA performed consistently better than both SPM and FAST for GM, WM, and CSF volume estimation. Compared with SPM and FAST, the proposed TRIOA showed more advantages by providing more accurate MR brain tissue classification and volume measurements, specifically in CSF volume estimation.

Brain Morphometry Estimation: From Hours to Seconds Using Deep Learning

Motivation: Brain morphometry from magnetic resonance imaging (MRI) is a promising neuroimaging biomarker for the non-invasive diagnosis and monitoring of neurodegenerative and neurological disorders. Current tools for brain morphometry often come with a high computational burden, making them hard to use in clinical routine, where time is often an issue. We propose a deep learning-based approach to predict the volumes of anatomically delineated subcortical regions of interest (ROI), and mean thicknesses and curvatures of cortical parcellations directly from T1-weighted MRI. Advantages are the timely availability of results while maintaining a clinically relevant accuracy. Materials and Methods: An anonymized dataset of 574 subjects (443 healthy controls and 131 patients with epilepsy) was used for the supervised training of a convolutional neural network (CNN). A silver-standard ground truth was generated with FreeSurfer 6.0. Results: The CNN predicts a total of 165 morphometric measures directly from raw MR images. Analysis of the results using intraclass correlation coefficients showed, in general, good correlation with FreeSurfer generated ground truth data, with some of the regions nearly reaching human inter-rater performance (ICC > 0.75). Cortical thicknesses predicted by the CNN showed cross-sectional annual age-related gray matter atrophy rates both globally (thickness change of -0.004 mm/year) and regionally in agreement with the literature. A statistical test to dichotomize patients with epilepsy from healthy controls revealed similar effect sizes for structures affecting all subtypes as reported in a large-scale epilepsy study. Conclusions: We demonstrate the general feasibility of using deep learning to estimate human brain morphometry directly from T1-weighted MRI within seconds. A comparison of the results to other publications shows accuracies of comparable magnitudes for the subcortical volumes and cortical thicknesses.