AI supported fetal echocardiography with quality assessment

This study aimed to develop a deep learning model to assess the quality of fetal echocardiography and to perform prospective clinical validation. The model was trained on data from the 18-22-week anomaly scan conducted in seven hospitals from 2008 to 2018. Prospective validation involved 100 patients from two hospitals. A total of 5363 images from 2551 pregnancies were used for training and validation. The model's segmentation accuracy depended on image quality measured by a quality score (QS). It achieved an overall average accuracy of 0.91 (SD 0.09) across the test set, with images having above-average QS scoring 0.97 (SD 0.03). During prospective validation of 192 images, clinicians rated 44.8% (SD 9.8) of images as equal in quality, 18.69% (SD 5.7) favoring auto-captured images and 36.51% (SD 9.0) preferring manually captured ones. Images with above average QS showed better agreement on segmentations (p < 0.001) and QS (p < 0.001) with fetal medicine experts. Auto-capture saved additional planes beyond protocol requirements, resulting in more comprehensive echocardiographies. Low QS had adverse effect on both model performance and clinician's agreement with model feedback. The findings highlight the importance of developing and evaluating AI models based on 'noisy' real-life data rather than pursuing the highest accuracy possible with retrospective academic-grade data.

Role of AI-assisted automated cardiac biometrics in screening for fetal coarctation of aorta

OBJECTIVES

Although there have been remarkable strides in fetal medicine and prenatal diagnosis of congenital heart disease, a significant percentage of newborns with isolated coarctation of the aorta (CoA) - around 60 percent - are still not identified prior to birth. The prenatal detection of CoA has been shown to have a notable impact on the survival rates of affected infants. To this end, the implementation of artificial intelligence (AI) in fetal ultrasound may represent a groundbreaking advancement. Our hypothesis is that leveraging automated cardiac biometric measurements with AI during the 18-22-week anomaly scan will enhance the identification of fetuses that are at risk of developing CoA.

METHODS

We have developed an AI model capable of identifying standard cardiac planes and conducting automated cardiac biometric measurements. Our data consisted of pregnancy ultrasound image and outcome data spanning from 2008 to 2018 and collected from four distinct regions in Denmark. The CoA cases from the period were paired with healthy controls in a ratio of 1:100 and matched on gestational ages of ±2 days. The cardiac biometrics on the four-chamber view and three vessel view were included in a logistic regression-based prediction model. To assess the predictive capabilities, we visualized sensitivity and specificity on Receiver Operating Characteristic (ROC) curves.

RESULTS

At the 18-22 week scan, the right ventricle (RV)area and length, left ventricle (LV) width, and the ratios of RV/LV areas and main pulmonary artery/ascending aorta diameters showed significant differences with z-scores above 0.7 when comparing subjects with a postnatal diagnosis of CoA (n=73) and healthy controls (n=7300). Using logistic regression and backward feature selection, our prediction model produced a ROC curve with an AUC (Area Under the Curve) of 0.96 and a specificity of 88.9% at a sensitivity level of 90.4%.

CONCLUSION

The integration of AI technology with automated cardiac biometric measurements conducted during the 18-22-week anomaly scan in fetal medicine has the potential to substantially enhance the screening for fetal CoA and subsequently the rate of CoA detection. Future research should clarify how AI technology can be used to aid in screening and detection of congenital heart anomalies to improve neonatal outcomes. This article is protected by copyright. All rights reserved.

The path toward equal performance in medical machine learning

To ensure equitable quality of care, differences in machine learning model performance between patient groups must be addressed. Here, we argue that two separate mechanisms can cause performance differences between groups. First, model performance may be worse than theoretically achievable in a given group. This can occur due to a combination of group underrepresentation, modeling choices, and the characteristics of the prediction task at hand. We examine scenarios in which underrepresentation leads to underperformance, scenarios in which it does not, and the differences between them. Second, the optimal achievable performance may also differ between groups due to differences in the intrinsic difficulty of the prediction task. We discuss several possible causes of such differences in task difficulty. In addition, challenges such as label biases and selection biases may confound both learning and performance evaluation. We highlight consequences for the path toward equal performance, and we emphasize that leveling up model performance may require gathering not only more data from underperforming groups but also better data. Throughout, we ground our discussion in real-world medical phenomena and case studies while also referencing relevant statistical theory.

Semantic similarity metrics for image registration

Image registration aims to find geometric transformations that align images. Most algorithmic and deep learning-based methods solve the registration problem by minimizing a loss function, consisting of a similarity metric comparing the aligned images, and a regularization term ensuring smoothness of the transformation. Existing similarity metrics like Euclidean Distance or Normalized Cross-Correlation focus on aligning pixel intensity values or correlations, giving difficulties with low intensity contrast, noise, and ambiguous matching. We propose a semantic similarity metric for image registration, focusing on aligning image areas based on semantic correspondence instead. Our approach learns dataset-specific features that drive the optimization of a learning-based registration model. We train both an unsupervised approach extracting features with an auto-encoder, and a semi-supervised approach using supplemental segmentation data. We validate the semantic similarity metric using both deep-learning-based and algorithmic image registration methods. Compared to existing methods across four different image modalities and applications, the method achieves consistently high registration accuracy and smooth transformation fields.

Populations of Unlabelled Networks: Graph Space Geometry and Generalized Geodesic Principal Components

Statistical analysis for populations of networks is widely applicable but challenging as networks have strongly non-Euclidean behaviour. Graph space is an exhaustive framework for studying populations of unlabelled networks which are weighted or unweighted, uni- or multi-layered, directed or undirected. Viewing graph space as the quotient of a Euclidean space with respect to a finite group action, we show that it is not a manifold, and that its curvature is unbounded from above. Within this geometrical framework we define generalized geodesic principal components, and we introduce the align all and compute algorithms, all of which allow for the computation of statistics on graph space. The statistics and algorithms are compared with existing methods and empirically validated on three real datasets, showcasing the framework potential utility. The whole framework is implemented within the geomstats Python package.

Multi-centre deep learning for placenta segmentation in obstetric ultrasound with multi-observer and cross-country generalization

The placenta is crucial to fetal well-being and it plays a significant role in the pathogenesis of hypertensive pregnancy disorders. Moreover, a timely diagnosis of placenta previa may save lives. Ultrasound is the primary imaging modality in pregnancy, but high-quality imaging depends on the access to equipment and staff, which is not possible in all settings. Convolutional neural networks may help standardize the acquisition of images for fetal diagnostics. Our aim was to develop a deep learning based model for classification and segmentation of the placenta in ultrasound images. We trained a model based on manual annotations of 7,500 ultrasound images to identify and segment the placenta. The model's performance was compared to annotations made by 25 clinicians (experts, trainees, midwives). The overall image classification accuracy was 81%. The average intersection over union score (IoU) reached 0.78. The model's accuracy was lower than experts' and trainees', but it outperformed all clinicians at delineating the placenta, IoU = 0.75 vs 0.69, 0.66, 0.59. The model was cross validated on 100 2nd trimester images from Barcelona, yielding an accuracy of 76%, IoU 0.68. In conclusion, we developed a model for automatic classification and segmentation of the placenta with consistent performance across different patient populations. It may be used for automated detection of placenta previa and enable future deep learning research in placental dysfunction.

Q-space trajectory imaging with positivity constraints (QTI+)

Q-space trajectory imaging (QTI) enables the estimation of useful scalar measures indicative of the local tissue structure. This is accomplished by employing generalized gradient waveforms for diffusion sensitization alongside a diffusion tensor distribution (DTD) model. The first two moments of the underlying DTD are made available by acquisitions at low diffusion sensitivity (b-values). Here, we show that three independent conditions have to be fulfilled by the mean and covariance tensors associated with distributions of symmetric positive semidefinite tensors. We introduce an estimation framework utilizing semi-definite programming (SDP) to guarantee that these conditions are met. Applying the framework on simulated signal profiles for diffusion tensors distributed according to non-central Wishart distributions demonstrates the improved noise resilience of QTI+ over the commonly employed estimation methods. Our findings on a human brain data set also reveal pronounced improvements, especially so for acquisition protocols featuring few number of volumes. Our method’s robustness to noise is expected to not only improve the accuracy of the estimates, but also enable a meaningful interpretation of contrast in the derived scalar maps. The technique’s performance on shorter acquisitions could make it feasible in routine clinical practice.

Enforcing necessary non-negativity constraints for common diffusion MRI models using sum of squares programming

In this work we investigate the use of sum of squares constraints for various diffusion-weighted MRI models, with a goal of enforcing strict, global non-negativity of the diffusion propagator. We formulate such constraints for the mean apparent propagator model and for spherical deconvolution, guaranteeing strict non-negativity of the corresponding diffusion propagators. For the cumulant expansion similar constraints cannot exist, and we instead derive a set of auxiliary constraints that are necessary but not sufficient to guarantee non-negativity. These constraints can all be verified and enforced at reasonable computational costs using semidefinite programming. By verifying our constraints on standard reconstructions of the different models, we show that currently used weak constraints are largely ineffective at ensuring non-negativity. We further show that if strict non-negativity is not enforced then estimated model parameters may suffer from significant errors, leading to serious inaccuracies in important derived quantities such as the main fiber orientations, mean kurtosis, etc. Finally, our experiments confirm that the observed constraint violations are mostly due to measurement noise, which is difficult to mitigate and suggests that properly constrained optimization should currently be considered the norm in many cases.

Scalable Robust Principal Component Analysis Using Grassmann Averages

In large datasets, manual data verification is impossible, and we must expect the number of outliers to increase with data size. While principal component analysis (PCA) can reduce data size, and scalable solutions exist, it is well-known that outliers can arbitrarily corrupt the results. Unfortunately, state-of-the-art approaches for robust PCA are not scalable. We note that in a zero-mean dataset, each observation spans a one-dimensional subspace, giving a point on the Grassmann manifold. We show that the average subspace corresponds to the leading principal component for Gaussian data. We provide a simple algorithm for computing this Grassmann Average ( GA), and show that the subspace estimate is less sensitive to outliers than PCA for general distributions. Because averages can be efficiently computed, we immediately gain scalability. We exploit robust averaging to formulate the Robust Grassmann Average (RGA) as a form of robust PCA. The resulting Trimmed Grassmann Average ( TGA) is appropriate for computer vision because it is robust to pixel outliers. The algorithm has linear computational complexity and minimal memory requirements. We demonstrate TGA for background modeling, video restoration, and shadow removal. We show scalability by performing robust PCA on the entire Star Wars IV movie; a task beyond any current method. Source code is available online.

Training shortest-path tractography: Automatic learning of spatial priors

Tractography is the standard tool for automatic delineation of white matter tracts from diffusion weighted images. However, the output of tractography often requires post-processing to remove false positives and ensure a robust delineation of the studied tract, and this demands expert prior knowledge. Here we demonstrate how such prior knowledge, or indeed any prior spatial information, can be automatically incorporated into a shortest-path tractography approach to produce more robust results. We describe how such a prior can be automatically generated (learned) from a population, and we demonstrate that our framework also retains support for conventional interactive constraints such as waypoint regions. We apply our approach to the open access, high quality Human Connectome Project data, as well as a dataset acquired on a typical clinical scanner. Our results show that the use of a learned prior substantially increases the overlap of tractography output with a reference atlas on both populations, and this is confirmed by visual inspection. Furthermore, we demonstrate how a prior learned on the high quality dataset significantly increases the overlap with the reference for the more typical yet lower quality data acquired on a clinical scanner. We hope that such automatic incorporation of prior knowledge and the obviation of expert interactive tract delineation on every subject, will improve the feasibility of large clinical tractography studies.

Geodesic Atlas-Based Labeling of Anatomical Trees: Application and Evaluation on Airways Extracted From CT

We present a fast and robust atlas-based algorithm for labeling airway trees, using geodesic distances in a geometric tree-space. Possible branch label configurations for an unlabeled airway tree are evaluated using distances to a training set of labeled airway trees. In tree-space, airway tree topology and geometry change continuously, giving a natural automatic handling of anatomical differences and noise. A hierarchical approach makes the algorithm efficient, assigning labels from the trachea and downwards. Only the airway centerline tree is used, which is relatively unaffected by pathology. The algorithm is evaluated on 80 segmented airway trees from 40 subjects at two time points, labeled by three medical experts each, testing accuracy, reproducibility and robustness in patients with chronic obstructive pulmonary disease (COPD). The accuracy of the algorithm is statistically similar to that of the experts and not significantly correlated with COPD severity. The reproducibility of the algorithm is significantly better than that of the experts, and negatively correlated with COPD severity. Evaluation of the algorithm on a longitudinal set of 8724 trees from a lung cancer screening trial shows that the algorithm can be used in large scale studies with high reproducibility, and that the negative correlation of reproducibility with COPD severity can be explained by missing branches, for instance due to segmentation problems in COPD patients. We conclude that the algorithm is robust to COPD severity given equally complete airway trees, and comparable in performance to that of experts in pulmonary medicine, emphasizing the suitability of the labeling algorithm for clinical use.

Toward a theory of statistical tree-shape analysis

To develop statistical methods for shapes with a tree-structure, we construct a shape space framework for tree-shapes and study metrics on the shape space. This shape space has singularities which correspond to topological transitions in the represented trees. We study two closely related metrics on the shape space, TED and QED. QED is a quotient euclidean distance arising naturally from the shape space formulation, while TED is the classical tree edit distance. Using Gromov's metric geometry, we gain new insight into the geometries defined by TED and QED. We show that the new metric QED has nice geometric properties that are needed for statistical analysis: Geodesics always exist and are generically locally unique. Following this, we can also show the existence and generic local uniqueness of average trees for QED. TED, while having some algorithmic advantages, does not share these advantages. Along with the theoretical framework we provide experimental proof-of-concept results on synthetic data trees as well as small airway trees from pulmonary CT scans. This way, we illustrate that our framework has promising theoretical and qualitative properties necessary to build a theory of statistical tree-shape analysis.

Tree-space statistics and approximations for large-scale analysis of anatomical trees

Statistical analysis of anatomical trees is hard to perform due to differences in the topological structure of the trees. In this paper we define statistical properties of leaf-labeled anatomical trees with geometric edge attributes by considering the anatomical trees as points in the geometric space of leaf-labeled trees. This tree-space is a geodesic metric space where any two trees are connected by a unique shortest path, which corresponds to a tree deformation. However, tree-space is not a manifold, and the usual strategy of performing statistical analysis in a tangent space and projecting onto tree-space is not available. Using tree-space and its shortest paths, a variety of statistical properties, such as mean, principal component, hypothesis testing and linear discriminant analysis can be defined. For some of these properties it is still an open problem how to compute them; others (like the mean) can be computed, but efficient alternatives are helpful in speeding up algorithms that use means iteratively, like hypothesis testing. In this paper, we take advantage of a very large dataset (N = 8016) to obtain computable approximations, under the assumption that the data trees parametrize the relevant parts of tree-space well. Using the developed approximate statistics, we illustrate how the structure and geometry of airway trees vary across a population and show that airway trees with Chronic Obstructive Pulmonary Disease come from a different distribution in tree-space than healthy ones. Software is available from http://image.diku.dk/aasa/software.php.