Accurate Segmentation of Vertebral Bodies and Processes Using Statistical Shape Decomposition and Conditional Models

LumVertCancNet: A novel 3D lumbar vertebral body cancellous bone location and segmentation method based on hybrid Swin-transformer

Lumbar vertebral body cancellous bone location and segmentation is crucial in an automated lumbar spine processing pipeline. Accurate and reliable analysis of lumbar spine image is expected to advantage practical medical diagnosis and population-based analysis of bone strength. However, the design of automated algorithms for lumbar spine processing is demanding due to significant anatomical variations and scarcity of publicly available data. In recent years, convolutional neural network (CNN) and vision transformers (Vits) have been the de facto standard in medical image segmentation. Although adept at capturing global features, the inherent bias of locality and weight sharing of CNN constrains its capacity to model long-range dependency. In contrast, Vits excel at long-range dependency modeling, but they may not generalize well with limited datasets due to the lack of inductive biases inherent to CNN. In this paper, we propose a deep learning-based two-stage coarse-to-fine solution to address the problem of automatic location and segmentation of lumbar vertebral body cancellous bone. Specifically, in the first stage, a Swin-transformer based model is applied to predict the heatmap of lumbar vertebral body centroids. Considering the characteristic anatomical structure of lumbar spine, we propose a novel loss function called LumAnatomy loss, which enforces the order and bend of the predicted vertebral body centroids. To inherit the excellence of CNN and Vits while preventing their respective limitations, in the second stage, we propose an encoder-decoder network to segment the identified lumbar vertebral body cancellous bone, which consists of two parallel encoders, i.e., a Swin-transformer encoder and a CNN encoder. To enhance the combination of CNNs and Vits, we propose a novel multi-scale attention feature fusion module (MSA-FFM), which address issues that arise when fusing features given at different encoders. To tackle the issue of lack of data, we raise the first large-scale lumbar vertebral body cancellous bone segmentation dataset called LumVBCanSeg containing a total of 185 CT scans annotated at voxel level by 3 physicians. Extensive experimental results on the LumVBCanSeg dataset demonstrate the proposed algorithm outperform other state-of-the-art medical image segmentation methods. The data is publicly available at: https://zenodo.org/record/8181250. The implementation of the proposed method is available at: https://github.com/sia405yd/LumVertCancNet.

An Automated Deep Learning Approach for Spine Segmentation and Vertebrae Recognition Using Computed Tomography Images

Spine image analysis is based on the accurate segmentation and vertebrae recognition of the spine. Several deep learning models have been proposed for spine segmentation and vertebrae recognition, but they are very computationally demanding. In this research, a novel deep learning model is introduced for spine segmentation and vertebrae recognition using CT images. The proposed model works in two steps: (1) A cascaded hierarchical atrous spatial pyramid pooling residual attention U-Net (CHASPPRAU-Net), which is a modified version of U-Net, is used for the segmentation of the spine. Cascaded spatial pyramid pooling layers, along with residual blocks, are used for feature extraction, while the attention module is used for focusing on regions of interest. (2) A 3D mobile residual U-Net (MRU-Net) is used for vertebrae recognition. MobileNetv2 includes residual and attention modules to accurately extract features from the axial, sagittal, and coronal views of 3D spine images. The features from these three views are concatenated to form a 3D feature map. After that, a 3D deep learning model is used for vertebrae recognition. The VerSe 20 and VerSe 19 datasets were used to validate the proposed model. The model achieved more accurate results in spine segmentation and vertebrae recognition than the state-of-the-art methods.

Prediction of the 3D shape of the L1 vertebral body from adjacent vertebrae

The aim of treatments of vertebral fractures is the anatomical reduction to restore the physiological biomechanics of the spine and the stabilization of the fracture to allow bone healing. However, the three-dimensional shape of the fractured vertebral body before the fracture is unknown in the clinical setting. Information about the pre-fracture vertebral body shape could help surgeons to select the optimal treatment. The goal of this study was to develop and validate a method based on Singular Value Decomposition (SVD) to predict the shape of the vertebral body of L1 from the shapes of T12 and L2. The geometry of the vertebral bodies of T12, L1 and L2 vertebrae of 40 patients were extracted from CT scans available from the VerSe2020 open-access dataset. Surface triangular meshes of each vertebra were morphed onto a template mesh. The set of vectors with the node coordinates of the morphed T12, L1 and L2 were compressed with SVD and used to build a system of linear equations. This system was used to solve a minimization problem and to reconstruct the shape of L1. A leave-one-out cross-validation was performed. Moreover, the approach was tested against an independent dataset with large osteophytes. The results of the study show a good prediction of the shape of the vertebral body of L1 from the shapes of the two adjacent vertebrae (mean error equal to 0.51 ± 0.11 mm on average, Hausdorff distance equal to 2.11 ± 0.56 mm on average), compared to current CT resolution typically used in the operating room. The error was slightly higher for patients presenting large osteophytes or severe bone degeneration (mean error equal to 0.65 ± 0.10 mm, Hausdorff distance equal to 3.54 ± 1.03 mm). The accuracy of the prediction was significantly better than approximating the shape of the vertebral body of L1 by the shape of T12 or L2. This approach could be used in the future to improve the pre-planning of spine surgeries to treat vertebral fractures.

CT-Based Automatic Spine Segmentation Using Patch-Based Deep Learning

CT vertebral segmentation plays an essential role in various clinical applications, such as computer-assisted surgical interventions, assessment of spinal abnormalities, and vertebral compression fractures. Automatic CT vertebral segmentation is challenging due to the overlapping shadows of thoracoabdominal structures such as the lungs, bony structures such as the ribs, and other issues such as ambiguous object borders, complicated spine architecture, patient variability, and fluctuations in image contrast. Deep learning is an emerging technique for disease diagnosis in the medical field. This study proposes a patch-based deep learning approach to extract the discriminative features from unlabeled data using a stacked sparse autoencoder (SSAE). 2D slices from a CT volume are divided into overlapping patches fed into the model for training. A random under sampling (RUS)-module is applied to balance the training data by selecting a subset of the majority class. SSAE uses pixel intensities alone to learn high-level features to recognize distinctive features from image patches. Each image is subjected to a sliding window operation to express image patches using autoencoder high-level features, which are then fed into a sigmoid layer to classify whether each patch is a vertebra or not. We validate our approach on three diverse publicly available datasets: VerSe, CSI-Seg, and the Lumbar CT dataset. Our proposed method outperformed other models after configuration optimization by achieving 89.9% in precision, 90.2% in recall, 98.9% in accuracy, 90.4% in F-score, 82.6% in intersection over union (IoU), and 90.2% in Dice coefficient (DC). The results of this study demonstrate that our model’s performance consistency using a variety of validation strategies is flexible, fast, and generalizable, making it suited for clinical application.

Morphable models of the lumbar spine to vary geometry based on pathology, demographics, and anatomical measurements

The shape of the lumbar spine influences its function and dysfunction. Yet examining the influence of geometric differences associated with pathology or demographics on lumbar biomechanics is challenging in vivo where these effects cannot be isolated, and the use of simple anatomical measurements does not fully capture the complex three-dimensional geometry. The goal of this work was to develop and share morphable models of the lumbar spine that allow geometry to be varied according to pathology, demographics, or anatomical measurements. Partial least squares regression was used to generate statistical shape models that quantify geometric differences associated with pathology, demographics, and anatomical measurements from the lumbar spines of 87 patients. To determine if the morphable models detected meaningful geometric differences, the ability of the morphable models to classify spines was compared with models generated from random labels. The models for disc herniation (p < 0.04), spondylolisthesis (p < 0.001), and sex (p < 0.01) all performed significantly better than the random models. Age was predicted with a root mean square error of 14.1 years using the age-based model. The morphable models for anatomical measurements were able to produce instances with root mean square errors less than 0.8°, 0.3 cm², and 0.7 mm between desired and resulting measurements. This method can be used to produce morphable models that enable further analysis of the relationship among shape, pathology, demographics, and function through computational simulations. The morphable models and code are available at https://github.com/aclouthier/morphable-lumbar-model.

Statistical Shape and Appearance Models: Development Towards Improved Osteoporosis Care

PURPOSE OF REVIEW

Statistical models of shape and appearance have increased their popularity since the 1990s and are today highly prevalent in the field of medical image analysis. In this article, we review the recent literature about how statistical models have been applied in the context of osteoporosis and fracture risk estimation.

RECENT FINDINGS

Recent developments have increased their ability to accurately segment bones, as well as to perform 3D reconstruction and classify bone anatomies, all features of high interest in the field of osteoporosis and fragility fractures diagnosis, prevention, and treatment. An increasing number of studies used statistical models to estimate fracture risk in retrospective case-control cohorts, which is a promising step towards future clinical application. All the reviewed application areas made considerable steps forward in the past 5-6 years. Heterogeneities in validation hinder a thorough comparison between the different methods and represent one of the future challenges to be addressed to reach clinical implementation.

The lumbar region localization using bone anatomy feature graphs

The automatic localization of the lumbar region is essential for the diagnosis of lumbar diseases, the study of lumbar morphology, and the surgical planning. Although the existing researches have made great progress, it still faces several challenges. First, the various lumbar diseases and pathologies cause different abnormalities in the lumbar shape and appearance. Second, the numbers of lumbar vertebrae are irregular (some people have an additional vertebra L6). To tackle these challenges, we propose a novel lumbar region localization method based on bone anatomy feature graphs. Specifically, a feature graph (called LS) considering the anatomy of the sacrum and the lumbar vertebra is proposed to locate the inferior boundary of L5 or L6. A feature graph (called TL) considering the anatomy of the thoracic vertebra and the lumbar vertebra is proposed to locate the superior boundary of L1. Extensive experimental analysis is performed on a public available dataset xVertSeg and a private dataset which contains 197 CT scans. The localization results show that the proposed method is robust and can be applied to normal scans, scoliosis scans, deformity scans, hyperosteogeny scans, 6 lumbar vertebrae scans and lumbar implant scans. The Dice and Jaccard coefficients are 98.09 ± 0.84% and 96.27 ± 1.62% respectively. Graphical Abstract Lumbar Region Localization Framework.

VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images

Vertebral labelling and segmentation are two fundamental tasks in an automated spine processing pipeline. Reliable and accurate processing of spine images is expected to benefit clinical decision support systems for diagnosis, surgery planning, and population-based analysis of spine and bone health. However, designing automated algorithms for spine processing is challenging predominantly due to considerable variations in anatomy and acquisition protocols and due to a severe shortage of publicly available data. Addressing these limitations, the Large Scale Vertebrae Segmentation Challenge (VerSe) was organised in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in 2019 and 2020, with a call for algorithms tackling the labelling and segmentation of vertebrae. Two datasets containing a total of 374 multi-detector CT scans from 355 patients were prepared and 4505 vertebrae have individually been annotated at voxel level by a human-machine hybrid algorithm (https://osf.io/nqjyw/, https://osf.io/t98fz/). A total of 25 algorithms were benchmarked on these datasets. In this work, we present the results of this evaluation and further investigate the performance variation at the vertebra level, scan level, and different fields of view. We also evaluate the generalisability of the approaches to an implicit domain shift in data by evaluating the top-performing algorithms of one challenge iteration on data from the other iteration. The principal takeaway from VerSe: the performance of an algorithm in labelling and segmenting a spine scan hinges on its ability to correctly identify vertebrae in cases of rare anatomical variations. The VerSe content and code can be accessed at: https://github.com/anjany/verse.

Anatomical fitting of a plate shape directly derived from a 3D statistical bone model of the tibia

INTRODUCTION

Intra- and inter-population variations of bone morphology have made the process of designing an anatomically well-fitting fracture fixation plate challenging. Although statistical bone models have recently been used for analysing morphological variabilities, it is not known to what extent they would also provide the basis for the design of a new plate shape. This would be particularly valuable in the case where no existing plate shape is available to start the process of fit optimisation. Therefore, this study investigated the anatomical fitting of a plate shape (statistical plate) derived from the mean shape of a statistical 3D tibia bone model in comparison to results available from two other plate shapes.

METHODS

Forty-five 3D bone models of tibiae from Japanese cadaver specimens, as well as 3D models of the plate undersurface of both a commercial and shape optimised Medial Distal Tibia Plate, were utilised from earlier studies. The mean shape of the 3D statistical bone model was generated from the tibia models utilising the Statismo framework. With reverse engineering software, the plate undersurface of the statistical plate shape was derived directly from the mean surface of the statistical 3D bone model. Through an iterative process, the statistical plate model was placed at the correct surgical position on each bone model for fit assessment.

RESULTS

The statistical plate was fitting for 20% of the tibiae compared to 13% for the commercial and 67% for the optimised plate, respectively.

CONCLUSIONS

The plate shape derived directly from a statistical bone model was fitting better than the commercial plate, but considerably inferior to that of an optimised plate. However, the results do clearly indicate that this approach provides an appropriate and solid basis for commencing shape optimisation of the statistical plate. Studies of other anatomical regions are required to confirm whether these findings can be generalised.

A novel robust kernel principal component analysis for nonlinear statistical shape modeling from erroneous data

Statistical Shape Models (SSMs) have achieved considerable success in medical image segmentation. A high quality SSM is able to approximate the main plausible variances of a given anatomical structure to guide segmentation. However, it is technically challenging to derive such a quality model because: (1) the distribution of shape variance is often nonlinear or multi-modal which cannot be modeled by standard approaches assuming Gaussian distribution; (2) as the quality of annotations in training data usually varies, heavy corruption will degrade the quality of the model as a whole. In this work, these challenges are addressed by introducing a generic SSM that is able to model nonlinear distribution and is robust to outliers in training data. Without losing generality and assuming a sparsity in nonlinear distribution, a novel Robust Kernel Principal Component Analysis (RKPCA) for statistical shape modeling is proposed with the aim of constructing a low-rank nonlinear subspace where outliers are discarded. The proposed approach is validated on two different datasets: a set of 30 public CT kidney pairs and a set of 49 MRI ankle bones volumes. Experimental results demonstrate a significantly better performance on outlier recovery and a higher quality of the proposed model as well as lower segmentation errors compared to the state-of-the-art techniques.

Lumbar model generator: a tool for the automated generation of a parametric scalable model of the lumbar spine

Low back pain is a major cause of disability and requires the development of new devices to treat pathologies and improve prognosis following surgery. Understanding the effects of new devices on the biomechanics of the spine is crucial in the development of new effective and functional devices. The aim of this study was to develop a preliminary parametric, scalable and anatomically accurate finite-element model of the lumbar spine allowing for the evaluation of the performance of spinal devices. The principal anatomical surfaces of the lumbar spine were first identified, and then accurately fitted from a previous model supplied by S14 Implants (Bordeaux, France). Finally, the reconstructed model was defined according to 17 parameters which are used to scale the model according to patient dimensions. The developed model, available as a toolbox named the lumbar model generator, enables generating a population of models using subject-specific dimensions obtained from data scans or averaged dimensions evaluated from the correlation analysis. This toolbox allows patient-specific assessment, taking into account individual morphological variation. The models have applications in the design process of new devices, evaluating the biomechanics of the spine and helping clinicians when deciding on treatment strategies.

Iterative fully convolutional neural networks for automatic vertebra segmentation and identification

Precise segmentation and anatomical identification of the vertebrae provides the basis for automatic analysis of the spine, such as detection of vertebral compression fractures or other abnormalities. Most dedicated spine CT and MR scans as well as scans of the chest, abdomen or neck cover only part of the spine. Segmentation and identification should therefore not rely on the visibility of certain vertebrae or a certain number of vertebrae. We propose an iterative instance segmentation approach that uses a fully convolutional neural network to segment and label vertebrae one after the other, independently of the number of visible vertebrae. This instance-by-instance segmentation is enabled by combining the network with a memory component that retains information about already segmented vertebrae. The network iteratively analyzes image patches, using information from both image and memory to search for the next vertebra. To efficiently traverse the image, we include the prior knowledge that the vertebrae are always located next to each other, which is used to follow the vertebral column. The network concurrently performs multiple tasks, which are segmentation of a vertebra, regression of its anatomical label and prediction whether the vertebra is completely visible in the image, which allows to exclude incompletely visible vertebrae from further analyses. The predicted anatomical labels of the individual vertebrae are additionally refined with a maximum likelihood approach, choosing the overall most likely labeling if all detected vertebrae are taken into account. This method was evaluated with five diverse datasets, including multiple modalities (CT and MR), various fields of view and coverages of different sections of the spine, and a particularly challenging set of low-dose chest CT scans. For vertebra segmentation, the average Dice score was 94.9 ± 2.1% with an average absolute symmetric surface distance of 0.2 ± 10.1mm. The anatomical identification had an accuracy of 93%, corresponding to a single case with mislabeled vertebrae. Vertebrae were classified as completely or incompletely visible with an accuracy of 97%. The proposed iterative segmentation method compares favorably with state-of-the-art methods and is fast, flexible and generalizable.

Automatic Deep Feature Learning via Patch-Based Deep Belief Network for Vertebrae Segmentation in CT Images

Precise automatic vertebra segmentation in computed tomography (CT) images is important for the quantitative analysis of vertebrae-related diseases but remains a challenging task due to high variation in spinal anatomy among patients. In this paper, we propose a deep learning approach for automatic CT vertebra segmentation named patch-based deep belief networks (PaDBNs). Our proposed PaDBN model automatically selects the features from image patches and then measures the differences between classes and investigates performance. The region of interest (ROI) is obtained from CT images. Unsupervised feature reduction contrastive divergence algorithm is applied for weight initialization, and the weights are optimized by layers in a supervised fine-tuning procedure. The discriminative learning features obtained from the steps above are used as input of a classifier to obtain the likelihood of the vertebrae. Experimental results demonstrate that the proposed PaDBN model can considerably reduce computational cost and produce an excellent performance in vertebra segmentation in terms of accuracy compared with state-of-the-art methods.

Method for Segmentation of Knee Articular Cartilages Based on Contrast-Enhanced CT Images

Segmentation of contrast-enhanced computed tomography (CECT) images enables quantitative evaluation of morphology of articular cartilage as well as the significance of the lesions. Unfortunately, automatic segmentation methods for CECT images are currently lacking. Here, we introduce a semiautomated technique to segment articular cartilage from in vivo CECT images of human knee. The segmented cartilage geometries of nine knee joints, imaged using a clinical CT-scanner with an intra-articular contrast agent, were compared with manual segmentations from CT and magnetic resonance (MR) images. The Dice similarity coefficients (DSCs) between semiautomatic and manual CT segmentations were 0.79-0.83 and sensitivity and specificity values were also high (0.76-0.86). When comparing semiautomatic and manual CT segmentations, mean cartilage thicknesses agreed well (intraclass correlation coefficient = 0.85-0.93); the difference in thickness (mean ± SD) was 0.27 ± 0.03 mm. Differences in DSC, when MR segmentations were compared with manual and semiautomated CT segmentations, were statistically insignificant. Similarly, differences in volume were not statistically significant between manual and semiautomatic CT segmentations. Semiautomation decreased the segmentation time from 450 ± 190 to 42 ± 10 min per joint. The results reveal that the proposed technique is fast and reliable for segmentation of cartilage. Importantly, this is the first study presenting semiautomated segmentation of cartilage from CECT images of human knee joint with minimal user interaction.

Nonrigid registration with corresponding points constraint for automatic segmentation of cardiac DSCT images

Background

Dual-source computed tomography (DSCT) is a very effective way for diagnosis and treatment of heart disease. The quantitative information of spatiotemporal DSCT images can be important for the evaluation of cardiac function. To avoid the shortcoming of manual delineation, it is imperative to develop an automatic segmentation technique for 4D cardiac images.

Methods

In this paper, we implement the heart segmentation-propagation framework based on nonrigid registration. The corresponding points of anatomical substructures are extracted by using the extension of n-dimensional scale invariant feature transform method. They are considered as a constraint term of nonrigid registration using the free-form deformation, in order to restrain the large variations and boundary ambiguity between subjects.

Results

We validate our method on 15 patients at ten time phases. Atlases are constructed by the training dataset from ten patients. On the remaining data the median overlap is shown to improve significantly compared to original mutual information, in particular from 0.4703 to 0.5015 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p = 5.0 \times 10^{ - 4} $$\end{document}p=5.0×10-4) for left ventricle myocardium and from 0.6307 to 0.6519 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p = 6.0 \times 10^{ - 4} $$\end{document}p=6.0×10-4) for right atrium.

Conclusions

The proposed method outperforms standard mutual information of intensity only. The segmentation errors had been significantly reduced at the left ventricle myocardium and the right atrium. The mean surface distance of using our framework is around 1.73 mm for the whole heart.

Computer-Assisted Screw Size and Insertion Trajectory Planning for Pedicle Screw Placement Surgery

Pathological conditions that cause instability of the spine are commonly treated by vertebral fixation involving pedicle screw placement surgery. However, existing methods for preoperative planning are based only on geometrical properties of vertebral structures (i.e., shape) without taking into account their structural properties (i.e., appearance). We propose a novel automated method for computer-assisted preoperative planning of the thoracic pedicle screw size and insertion trajectory. The proposed method extracts geometrical properties of vertebral structures by parametric modeling of vertebral bodies and pedicles in three dimensions (3D), and combines them with structural properties, evaluated through underlying image intensities in computed tomography (CT) images while considering the guidelines for pedicle screw design. The method was evaluated on 81 pedicles, obtained from 3D CT images of 11 patients that were appointed for pedicle screw placement surgery. In terms of mean absolute difference (MAD) and corresponding standard deviation (SD), the resulting high modeling accuracy of 0.39±0.31 mm for 3D vertebral body models and 0.31±0.25 mm for 3D pedicle models created an adequate anatomical frame for 3D pedicle screw models. When comparing the automatically obtained and manually defined plans for pedicle screw placement, a relatively high agreement was observed, with MAD ±SD of 0.4±0.4 mm for the screw diameter, 5.8±4.2 mm for the screw length, 2.0±1.4 mm for the pedicle crossing point and 7.6±5.8(°) for screw insertion angles. However, a statistically significant increase of 48±26% in the screw fastening strength in favor of the proposed automated method was observed in 99% of the cases.