Full Segmentations Research Articles

Segmentation evaluation with sparse ground truth data: Simulating true segmentations as perfect/imperfect as those generated by humans.

Fully annotated data sets play important roles in medical image segmentation and evaluation. Expense and imprecision are the two main issues in generating ground truth (GT) segmentations. In this paper, in an attempt to overcome these two issues jointly, we propose a method, named SparseGT, which exploit variability among human segmenters to maximally save manual workload in GT generation for evaluating actual segmentations by algorithms. Pseudo ground truth (p-GT) segmentations are created by only a small fraction of workload and with human-level perfection/imperfection, and they can be used in practice as a substitute for fully manual GT in evaluating segmentation algorithms at the same precision. p-GT segmentations are generated by first selecting slices sparsely, where manual contouring is conducted only on these sparse slices, and subsequently filling segmentations on other slices automatically. By creating p-GT with different levels of sparseness, we determine the largest workload reduction achievable for each considered object, where the variability of the generated p-GT is statistically indistinguishable from inter-segmenter differences in full manual GT segmentations for that object. Furthermore, we investigate the segmentation evaluation errors introduced by variability in manual GT by applying p-GT in evaluation of actual segmentations by an algorithm. Experiments are conducted on ∼500 computed tomography (CT) studies involving six objects in two body regions, Head & Neck and Thorax, where optimal sparseness and corresponding evaluation errors are determined for each object and each strategy. Our results indicate that creating p-GT by the concatenated strategy of uniformly selecting sparse slices and filling segmentations via deep-learning (DL) network show highest manual workload reduction by ∼80-96% without sacrificing evaluation accuracy compared to fully manual GT. Nevertheless, other strategies also have obvious contributions in different situations. A non-uniform strategy for slice selection shows its advantage for objects with irregular shape change from slice to slice. An interpolation strategy for filling segmentations can achieve ∼60-90% of workload reduction in simulating human-level GT without the need of an actual training stage and shows potential in enlarging data sets for training p-GT generation networks. We conclude that not only over 90% reduction in workload is feasible without sacrificing evaluation accuracy but also the suitable strategy and the optimal sparseness level achievable for creating p-GT are object- and application-specific.

Development and evaluation of a manual segmentation protocol for deep grey matter in multiple sclerosis: Towards accelerated semi-automated references

BackgroundDeep grey matter (dGM) structures, particularly the thalamus, are clinically relevant in multiple sclerosis (MS). However, segmentation of dGM in MS is challenging; labeled MS-specific reference sets are needed for objective evaluation and training of new methods. ObjectivesThis study aimed to (i) create a standardized protocol for manual delineations of dGM; (ii) evaluate the reliability of the protocol with multiple raters; and (iii) evaluate the accuracy of a fast-semi-automated segmentation approach (FASTSURF). MethodsA standardized manual segmentation protocol for caudate nucleus, putamen, and thalamus was created, and applied by three raters on multi-center 3D T1-weighted MRI scans of 23 MS patients and 12 controls. Intra- and inter-rater agreement was assessed through intra-class correlation coefficient (ICC); spatial overlap through Jaccard Index (JI) and generalized conformity index (CIgen). From sparse delineations, FASTSURF reconstructed full segmentations; accuracy was assessed both volumetrically and spatially. ResultsAll structures showed excellent agreement on expert manual outlines: intra-rater JI > 0.83; inter-rater ICC ≥ 0.76 and CIgen ≥ 0.74. FASTSURF reproduced manual references excellently, with ICC ≥ 0.97 and JI ≥ 0.92. ConclusionsThe manual dGM segmentation protocol showed excellent reproducibility within and between raters. Moreover, combined with FASTSURF a reliable reference set of dGM segmentations can be produced with lower workload.

Muscle and Tendon Morphology in Early-Adolescent Athletes and Untrained Peers.

Adolescent athletes can feature significantly greater muscle strength and tendon stiffness compared to untrained peers. However, to date, it is widely unclear if radial muscle and tendon hypertrophy may contribute to loading-induced adaptation at this stage of maturation. The present study compares the morphology of the vastus lateralis (VL) and the patellar tendon between early-adolescent athletes and untrained peers. In 14 male elite athletes (A) and 10 untrained controls (UC; 12–14 years of age), the VL was reconstructed from full muscle segmentations of magnetic resonance imaging (MRI) sequences and ultrasound imaging was used to measure VL fascicle length and pennation angle. The physiological cross-sectional area (PCSA) of the VL was calculated by dividing muscle volume by fascicle length. The cross-sectional area (CSA) of the patellar tendon was measured over its length based on MRI segmentations as well. Considering body mass as covariate in the analysis, there were no significant differences between groups considering the VL anatomical cross-sectional area (ACSA) over its length or maximum ACSA (UC: 24.0 ± 8.3 cm2, A: 28.1 ± 5.3 cm2, p > 0.05), yet athletes had significantly greater VL volume (UC: 440 ± 147 cm3, A: 589 ± 121 cm3), PCSA (UC: 31 ± 9 cm2, A: 46 ± 9 cm2), pennation angle (UC: 8.2 ± 1.4°, A: 10.1 ± 1.3°), and average patellar tendon CSA (UC: 1.01 ± 0.18 cm2, A: 1.21 ± 0.18 cm2) compared to the untrained peers (p < 0.05). However, the ratio of average tendon CSA to VL PCSA was significantly lower in athletes (UC: 3.4 ± 0.1%, A: 2.7 ± 0.5%; p < 0.05). When inferring effects of athletic training based on the observed differences between groups, these results suggest that both muscle and tendon of the knee extensors respond to athletic training with radial growth. However, the effect seems to be stronger in the muscle compared to the tendon, with an increase of pennation angle contributing to the marked increase of muscle PCSA. A disproportionate response to athletic training might be associated with imbalances of muscle strength and tendon stiffness and could have implications for the disposition towards tendon overuse injury.

Development and validation of statistical shape models of the primary functional bone segments of the foot.

IntroductionMusculoskeletal models are important tools for studying movement patterns, tissue loading, and neuromechanics. Personalising bone anatomy within models improves analysis accuracy. Few studies have focused on personalising foot bone anatomy, potentially incorrectly estimating the foot’s contribution to locomotion. Statistical shape models have been created for a subset of foot-ankle bones, but have not been validated. This study aimed to develop and validate statistical shape models of the functional segments in the foot: first metatarsal, midfoot (second-to-fifth metatarsals, cuneiforms, cuboid, and navicular), calcaneus, and talus; then, to assess reconstruction accuracy of these shape models using sparse anatomical data.MethodsMagnetic resonance images of 24 individuals feet (age = 28 ± 6 years, 52% female, height = 1.73 ± 0.8 m, mass = 66.6 ± 13.8 kg) were manually segmented to generate three-dimensional point clouds. Point clouds were registered and analysed using principal component analysis. For each bone segment, a statistical shape model and principal components were created, describing population shape variation. Statistical shape models were validated by assessing reconstruction accuracy in a leave-one-out cross validation. Statistical shape models were created by excluding a participant’s bone segment and used to reconstruct that same excluded bone using full segmentations and sparse anatomical data (i.e. three discrete points on each segment), for all combinations in the dataset. Tali were not reconstructed using sparse anatomical data due to a lack of externally accessible landmarks. Reconstruction accuracy was assessed using Jaccard index, root mean square error (mm), and Hausdorff distance (mm).ResultsReconstructions generated using full segmentations had mean Jaccard indices between 0.77 ± 0.04 and 0.89 ± 0.02, mean root mean square errors between 0.88 ± 0.19 and 1.17 ± 0.18 mm, and mean Hausdorff distances between 2.99 ± 0.98 mm and 6.63 ± 3.68 mm. Reconstructions generated using sparse anatomical data had mean Jaccard indices between 0.67 ± 0.06 and 0.83 ± 0.05, mean root mean square error between 1.21 ± 0.54 mm and 1.66 ± 0.41 mm, and mean Hausdorff distances between 3.21 ± 0.94 mm and 7.19 ± 3.54 mm. Jaccard index was higher (P < 0.01) and root mean square error was lower (P < 0.01) in reconstructions from full segmentations compared to sparse anatomical data. Hausdorff distance was lower (P < 0.01) for midfoot and calcaneus reconstructions using full segmentations compared to sparse anatomical data.ConclusionFor the first time, statistical shape models of the primary functional segments of the foot were developed and validated. Foot segments can be reconstructed with minimal error using full segmentations and sparse anatomical landmarks. In future, larger training datasets could increase statistical shape model robustness, extending use to paediatric or pathological populations.

Machine deep learning accurately detects endoleak after endovascular abdominal aortic aneurysm repair

ObjectiveThe objective of this study was to develop a machine deep learning algorithm for endoleak detection and measurement of aneurysm diameter, area, and volume from computed tomography angiography (CTA). MethodsDigital Imaging and Communications in Medicine files representing three-phase postoperative CTA images (N = 334) of 191 unique patients undergoing endovascular aneurysm repair for infrarenal abdominal aortic aneurysm (AAA) with a variety of commercial devices were used to train a deep learning pipeline across four tasks. The RetinaNet object-detection convolutional neural network (CNN) architecture was trained to predict bounding boxes around the axial CTA slices that were then stitched together in two dimensions into a smaller region containing the aneurysm. Multiclass endoleak detection and segmentation of the AAA, endograft, and endoleak were performed on this smaller region. Segmentations on a single randomly selected contrast from each scan included 33 full and 68 partial segmentations for endograft and AAA and 99 full segmentations for endoleak. A modified version of ResNet-50 CNN was used to detect endoleak on individual axial slices. A three-dimensional U-Net CNN model was trained on the task of dense three-dimensional segmentation and used to measure diameter and volume with a specially designed loss function. We made use of fivefold cross-validation to evaluate model performance for each step, splitting training and testing data at each fold, such that multiple scans from the same patient were preserved with the same fold. Algorithm predictions for endoleak were compared with the radiology report and with a subset of CTA images independently read by two vascular specialists. ResultsThe localization portion of the network accurately predicted a region of interest containing the AAA in 99% of cases. The best model of binary endoleak detection obtained an area under the receiver operating characteristic curve of 0.94 ± 0.03 with an optimized accuracy of 0.89 ± 0.03 on a balanced data set. An introduced postprocessing algorithm for determining maximum diameter was used on both the predicted AAA segmentation and ground truth segmentation, predicting on average an absolute diameter error of 2.3 ± 2.0 mm by 1.4 ± 1.7 mm for each measurement, respectively. The algorithm measured AAA and endograft volume accurately (Dice coefficient, 0.95 ± 0.2) with an absolute volume error of 10.1 ± 9.1 mL. The algorithm measured endoleak volume less accurately, with a Dice score of 0.53 ± 0.21 and an average absolute volume error of 1.2 ± 1.9 mL. ConclusionsThis machine learning algorithm shows promise in augmenting a human's ability to interpret postoperative CTA images and may help improve surveillance after endovascular aneurysm repair. External validation on larger data sets and prospective study are required before the algorithm can be clinically applicable.

FAst Segmentation Through SURface Fairing (FASTSURF): A novel semi-automatic hippocampus segmentation method.

ObjectiveThe objective is to present a proof-of-concept of a semi-automatic method to reduce hippocampus segmentation time on magnetic resonance images (MRI).Materials and methodsFAst Segmentation Through SURface Fairing (FASTSURF) is based on a surface fairing technique which reconstructs the hippocampus from sparse delineations. To validate FASTSURF, simulations were performed in which sparse delineations extracted from full manual segmentations served as input. On three different datasets with different diagnostic groups, FASTSURF hippocampi were compared to the original segmentations using Jaccard overlap indices and percentage volume differences (PVD). In one data set for which back-to-back scans were available, unbiased estimates of overlap and PVD were obtained. Using longitudinal scans, we compared hippocampal atrophy rates measured by manual, FASTSURF and two automatic segmentations (FreeSurfer and FSL-FIRST).ResultsWith only seven input contours, FASTSURF yielded mean Jaccard indices ranging from 72(±4.3)% to 83(±2.6)% and PVDs ranging from 0.02(±2.40)% to 3.2(±3.40)% across the three datasets. Slightly poorer results were obtained for the unbiased analysis, but the performance was still considerably better than both tested automatic methods with only five contours.ConclusionsFASTSURF segmentations have high accuracy and require only a fraction of the delineation effort of fully manual segmentation. Atrophy rate quantification based on completely manual segmentation is well reproduced by FASTSURF. Therefore, FASTSURF is a promising tool to be implemented in clinical workflow, provided a future prospective validation confirms our findings.

Hemodynamic Changes in the Treatment of Multiple Intracranial Aneurysms: A Computational Fluid Dynamics Study

Intracranial aneurysm rupture is associated with a high mortality and disability despite modern medical care. Multiple aneurysms occur in nearly 16% of patients, and imaging studies of naturally occurring multiple aneurysms are valuable for computational fluid dynamics studies. In this study, we describe and analyze the hemodynamic changes produced in a distal aneurysm after the treatment of a proximal aneurysm. Between January 2008 and December 2017, 24 cases of multiple intracranial aneurysms of the same vascular tree were treated in our center. Full carotid segmentations from digital subtraction angiography, computed tomography angiography, or magnetic resonance angiography were obtained, and transient pulsatile simulations were performed using computational fluid dynamics software. Output variables included maximum peak systole wall shear stress (WSS), minimum mid-diastolic WSS, maximum systolic pressure, low shear area, and high shear area both before and after treatment of the proximal aneurysm. The mean size of ruptured and unruptured aneurysms was 7.05 and 5.23 mm, respectively (P= 0.035), with respective aspect ratios of 1.22 and 2.04 (P= 0.001). Maximum peak systole WSS was lower and minimum mid-diastolic WSS was higher in unruptured aneurysms (P=0.04 and 0.034, respectively). After treatment of the proximal aneurysm, low shear area in the distal aneurysm increased from 54.15% to 56.93% (P= 0.02). The opposite effect is noted in aneurysms with an interaneurysmal distance <10 mm. Posttreatment peak systole pressure was also increased significantly (P= 0.03). The hemodynamic changes in a distal aneurysm after treatment of a proximal aneurysm showed an unfavorable profile associated with an increased theoretical risk of bleeding.

Multiple RF classifier for the hippocampus segmentation: Method and validation on EADC-ADNI Harmonized Hippocampal Protocol

The hippocampus has a key role in a number of neurodegenerative diseases, such as Alzheimer's Disease. Here we present a novel method for the automated segmentation of the hippocampus from structural magnetic resonance images (MRI), based on a combination of multiple classifiers. The method is validated on a cohort of 50 T1 MRI scans, comprehending healthy control, mild cognitive impairment, and Alzheimer's Disease subjects. The preliminary release of the EADC-ADNI Harmonized Protocol training labels is used as gold standard. The fully automated pipeline consists of a registration using an affine transformation, the extraction of a local bounding box, and the classification of each voxel in two classes (background and hippocampus). The classification is performed slice-by-slice along each of the three orthogonal directions of the 3D-MRI using a Random Forest (RF) classifier, followed by a fusion of the three full segmentations. Dice coefficients obtained by multiple RF (0.87 ± 0.03) are larger than those obtained by a single monolithic RF applied to the entire bounding box, and are comparable to state-of-the-art. A test on an external cohort of 50 T1 MRI scans shows that the presented method is robust and reliable. Additionally, a comparison of local changes in the morphology of the hippocampi between the three subject groups is performed. Our work showed that a multiple classification approach can be implemented for the segmentation for the measurement of volume and shape changes of the hippocampus with diagnostic purposes.

Biomechanically Constrained Surface Registration: Application to MR-TRUS Fusion for Prostate Interventions.

In surface-based registration for image-guided interventions, the presence of missing data can be a significant issue. This often arises with real-time imaging modalities such as ultrasound, where poor contrast can make tissue boundaries difficult to distinguish from surrounding tissue. Missing data poses two challenges: ambiguity in establishing correspondences; and extrapolation of the deformation field to those missing regions. To address these, we present a novel non-rigid registration method. For establishing correspondences, we use a probabilistic framework based on a Gaussian mixture model (GMM) that treats one surface as a potentially partial observation. To extrapolate and constrain the deformation field, we incorporate biomechanical prior knowledge in the form of a finite element model (FEM). We validate the algorithm, referred to as GMM-FEM, in the context of prostate interventions. Our method leads to a significant reduction in target registration error (TRE) compared to similar state-of-the-art registration algorithms in the case of missing data up to 30%, with a mean TRE of 2.6 mm. The method also performs well when full segmentations are available, leading to TREs that are comparable to or better than other surface-based techniques. We also analyze robustness of our approach, showing that GMM-FEM is a practical and reliable solution for surface-based registration.

Materials-based 3D segmentation of unknown objects from dual-energy computed tomography imagery in baggage security screening

We present a novel technique for the 3D segmentation of unknown objects from cluttered dual-energy Computed Tomography (CT) data obtained in the baggage security-screening domain. Initial materials-based coarse segmentations, generated using the Dual-Energy Index (DEI), are refined by partitioning at automatically detected regions. Partitioning is guided by a novel random forest based quality metric, trained to recognise high-quality, single-object segments. A second novel segmentation quality measure is presented for quantifying the quality of full segmentations based on the random forest metric of the constituent parts and the error in the number of objects segmented. In a comparative evaluation between the proposed approach and three state-of-the-art volumetric segmentation techniques designed for single-energy CT data (two region-growing [1,2] and one graph-based [3]) our method is shown to outperform both region-growing methods in terms of segmentation quality and speed. Although the graph-based approach generates more accurate partitions, it is characterised by high processing times and is significantly outperformed by the proposed method in this regard. The observations made in this study indicate that the proposed segmentation technique is well-suited to the baggage security-screening domain, where the demand for computational efficiency is paramount to maximise throughput.

Investigation of magnet segmentation techniques for eddy current losses reduction in permanent magnets electrical machines

Purpose – Reducing eddy current losses in magnets of electrical machines can be obtained by means of several techniques. The magnet segmentation is the most popular one. It imposes the least restrictions on machine performances. This paper investigates the effectiveness of the magnet circumferential segmentation technique to reduce these undesirable losses. The full and partial magnet segmentation are both studied for a frequency range from few Hz to a dozen of kHz. To increase the efficiency of these techniques to reduce losses for any working frequency, an optimization strategy based on coupling of finite elements analysis and genetic algorithm is applied. The purpose of this paper is to define the parameters of the total and partial segmentation that can ensure the best reduction of eddy current losses. Design/methodology/approach – First, a model to analyze eddy current losses is presented. Second, the effectiveness of full and partial magnet circumferential segmentation to reduce eddy loss is studied for a range of frequencies from few Hz to a dozen of kHz. To achieve these purposes a 2-D finite element model is developed under MATLAB environment. In a third step of the work, an optimization process is applied to adjust the segmentation design parameters for best reduction of eddy current losses in case of surface mounted permanent magnets synchronous machine. Findings – In case of the skin effect operating, both full and partial magnet segmentations can lead to eddy current losses increases. Such deviations of magnet segmentation techniques can be avoided by an appropriate choice of their design parameters. Originality/value – Few works are dedicated to investigate partial magnet segmentation for eddy current losses reduction. This paper studied the effectiveness and behaviour of partial segmentation for different frequency ranges. To avoid eventual anomalies related to the skin effect an optimization process based on the association of the finite elements analysis to genetic algorithm method is adopted.

MRBrainS Challenge: Online Evaluation Framework for Brain Image Segmentation in 3T MRI Scans

Many methods have been proposed for tissue segmentation in brain MRI scans. The multitude of methods proposed complicates the choice of one method above others. We have therefore established the MRBrainS online evaluation framework for evaluating (semi)automatic algorithms that segment gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) on 3T brain MRI scans of elderly subjects (65–80 y). Participants apply their algorithms to the provided data, after which their results are evaluated and ranked. Full manual segmentations of GM, WM, and CSF are available for all scans and used as the reference standard. Five datasets are provided for training and fifteen for testing. The evaluated methods are ranked based on their overall performance to segment GM, WM, and CSF and evaluated using three evaluation metrics (Dice, H95, and AVD) and the results are published on the MRBrainS13 website. We present the results of eleven segmentation algorithms that participated in the MRBrainS13 challenge workshop at MICCAI, where the framework was launched, and three commonly used freeware packages: FreeSurfer, FSL, and SPM. The MRBrainS evaluation framework provides an objective and direct comparison of all evaluated algorithms and can aid in selecting the best performing method for the segmentation goal at hand.

3D of Brain Shape and Volume After Cranial Vault Remodeling Surgery for Craniosynostosis Correction in Infants.

The skull of young children is made up of bony plates that enable growth. Craniosynostosis is a birth defect that causes one or more sutures on an infant's skull to close prematurely. Corrective surgery focuses on cranial and orbital rim shaping to return the skull to a more normal shape. Functional problems caused by craniosynostosis such as speech and motor delay can improve after surgical correction, but a post-surgical analysis of brain development in comparison with age-matched healthy controls is necessary to assess surgical outcome. Full brain segmentations obtained from pre- and post-operative computed tomography (CT) scans of 8 patients with single suture sagittal (n=5) and metopic (n=3), non-syndromic craniosynostosis from 41 to 452 days-of-age were included in this study. Age-matched controls obtained via 4D acceleration-based regression of a cohort of 402 full brain segmentations from healthy controls magnetic resonance images (MRI) were also used for comparison (ages 38 to 825 days). 3D point-based models of patient and control cohorts were obtained using SPHARM-PDM shape analysis tool. From a full dataset of regressed shapes, 240 healthy regressed shapes between 30 and 588 days-of-age (time step = 2.34 days) were selected. Volumes and shape metrics were obtained for craniosynostosis and healthy age-matched subjects. Volumes and shape metrics in single suture craniosynostosis patients were larger than age-matched controls for pre- and post-surgery. The use of 3D shape and volumetric measurements show that brain growth is not normal in patients with single suture craniosynostosis.