Abstract:

Introduction: Photogrammetric surface scans provide a radiation-free option to assess and classify craniosynostosis. Due to the low prevalence of craniosynostosis and high patient restrictions, clinical data is rare. Synthetic data could support or even replace clinical data for the classification of craniosynostosis, but this has never been studied systematically. Methods: We test the combinations of three different synthetic data sources: a statistical shape model (SSM), a generative adversarial network (GAN), and image-based principal component analysis for a convolutional neural network (CNN)-based classification of craniosynostosis. The CNN is trained only on synthetic data, but validated and tested on clinical data. Results: The combination of a SSM and a GAN achieved an accuracy of more than 0.96 and a F1-score of more than 0.95 on the unseen test set. The difference to training on clinical data was smaller than 0.01. Including a second image modality improved classification performance for all data sources. Conclusion: Without a single clinical training sample, a CNN was able to classify head deformities as accurate as if it was trained on clinical data. Using multiple data sources was key for a good classification based on synthetic data alone. Synthetic data might play an important future role in the assessment of craniosynostosis.

Abstract:

Objective: Diagnosis of craniosynostosis using photogrammetric 3D surface scans is a promising radiation-free alternative to traditional computed tomography. We propose a 3D surface scan to 2D distance map conversion enabling the usage of the first convolutional neural networks (CNNs)-based classification of craniosynostosis. Benefits of using 2D images include preserving patient anonymity, enabling data augmentation during training, and a strong under-sampling of the 3D surface with good classification performance.Methods: The proposed distance maps sample 2D images from 3D surface scans using a coordinate transformation, ray casting, and distance extraction. We introduce a CNNbased classification pipeline and compare our classifier to alternative approaches on a dataset of 496 patients. We investigate into low-resolution sampling, data augmentation, and attribution mapping.Results: Resnet18 outperformed alternative classifiers on our dataset with an F1-score of 0.964 and an accuracy of 98.4 %. Data augmentation on 2D distance maps increased performance for all classifiers. Under-sampling allowed 256-fold computation reduction during ray casting while retaining an F1-score of 0.92. Attribution maps showed high amplitudes on the frontal head.Conclusion: We demonstrated a versatile mapping approach to extract a 2D distance map from the 3D head geometry increasing classification performance, enabling data augmentation during training on 2D distance maps, and the usage of CNNs. We found that low-resolution images were sufficient for a good classification performance.Significance: Photogrammetric surface scans are a suitable craniosynostosis diagnosis tool for clinical practice. Domain transfer to computed tomography seems likely and can further contribute to reducing ionizing radiation exposure for infants.

Abstract:

Background: Craniosynostosis is a condition caused by the premature fusion of skull sutures, leading to irregular growth patterns of the head. Three-dimensional photogrammetry is a radiation-free alternative to the diagnosis using computed tomography. While statistical shape models have been proposed to quantify head shape, no shape-model-based classification approach has been presented yet. Methods: We present a classification pipeline that enables an automated diagnosis of three types of craniosynostosis. The pipeline is based on a statistical shape model built from photogrammetric surface scans. We made the model and pathology-specific submodels publicly available, making it the first publicly available craniosynostosis-related head model, as well as the first focusing on infants younger than 1.5 years. To the best of our knowledge, we performed the largest classification study for craniosynostosis to date. Results: Our classification approach yields an accuracy of 97.8 %, comparable to other state-of-the-art methods using both computed tomography scans and stereophotogrammetry. Regarding the statistical shape model, we demonstrate that our model performs similar to other statistical shape models of the human head. Conclusion: We present a state-of-the-art shape-model-based classification approach for a radiation-free diagnosis of craniosynostosis. Our publicly available shape model enables the assessment of craniosynostosis on realistic and synthetic data.

Abstract:

This contribution is part of a project concerning the creation of an artificial dataset comprising 3D head scans of craniosynostosis patients for a deep-learning-based classification. To conform to real data, both head and neck are required in the 3D scans. However, during patient recording, the neck is often covered by medical staff. Simply pasting an arbitrary neck leaves large gaps in the 3D mesh. We therefore use a publicly available statistical shape model (SSM) for neck reconstruction. However, most SSMs of the head are constructed using healthy subjects, so the full head reconstruction loses the craniosynostosis-specific head shape. We propose a method to recover the neck while keeping the pathological head shape intact. We propose a Laplace- Beltrami-based refinement step to deform the posterior mean shape of the full head model towards the pathological head. The artificial neck is created using the publicly available Liverpool-Y ork-Model. W e apply our method to construct artificial necks for head scans of 50 scaphocephaly patients. Our method reduces mean vertex correspondence error by approximately 1.3 mm compared to the ordinary posterior mean shape, preserves the pathological head shape, and creates a continuous transition between neck and head. The presented method showed good results for reconstructing a plausible neck to craniosynostosis patients. Easily generalized it might also be applicable to other pathological shapes.

Abstract:

The application of machine learning approachesin medical technology is gaining more and more attention.Due to the high restrictions for collecting intraoperative patientdata, synthetic data is increasingly used to support the trainingof artificial neural networks. We present a pipeline to createa statistical shape model (SSM) using 28 segmented clinicalliver CT scans. Our pipeline consists of four steps: data pre-processing, rigid alignment, template morphing, and statisti-cal modeling. We compared two different template morphingapproaches: Laplace-Beltrami-regularized projection (LBRP)and nonrigid iterative closest points translational (N-ICP-T)and evaluated both morphing approaches and their corre-sponding shape model performance using six metrics. LBRPachieved a smaller mean vertex-to-nearest-neighbor distances(2.486±0.897 mm) than N-ICP-T (5.559±2.413 mm). Gen-eralizationand specificity errors for LBRP were consistentlylower than those of N-ICP-T. The first principal componentsof the SSM showed realistic anatomical variations. The perfor-mance of the SSM was comparable to a state-of-the-art model.

Abstract:

Introduction: 3D surface scan-based diagnosis of craniosynostosis is a promising radiation-free alternative to traditional diagnosis using computed tomography. The cra- nial index (CI) and the cranial vault asymmetry index (CVAI) are well-established clinical parameters that are widely used. However, they also have the benefit of being easily adaptable for automatic diagnosis without the need of extensive prepro- cessing.Methods: We propose a multi-height-based classification ap- proach that uses CI and CVAI in different height layers and compare it to the initial approach using only one layer. We use ten-fold cross-validation and test seven different classi- fiers. The dataset of 504 patients consists of three types of craniosynostosis and a control group consisting of healthy and non-synostotic subjects.Results: The multi-height-based approach improved classifica- tion for all classifiers. The k-nearest neighbors classifier scored best with a mean accuracy of 89 % and a mean F1-score of 0.75.Conclusion: Taking height into account is beneficial for the classification. Based on accepted and widely used clinical pa- rameters, this might be a step towards an easy-to-understand and transparent classification approach for both physicians and patients.

Abstract:

Craniosynostosis is a congenital disease character-ized by the premature closure of one or multiple sutures of theinfant’s skull. For diagnosis, 3D photogrammetric scans are aradiation-free alternative to computed tomography. However,data is only sparsely available and the role of data augmentation for the classification of craniosynostosis has not yet beenanalyzed.In this work, we use a 2D distance map representation ofthe infants’ heads with a convolutional-neural-network-basedclassifier and employ a generative adversarial network (GAN)for data augmentation. We simulate two data scarcity scenar-ios with 15 % and 10 % training data and test the influence ofdifferent degrees of added synthetic data and balancing under-represented classes. We used total accuracy and F1-score as ametric to evaluate the final classifiers.For 15 % training data, the GAN-augmented dataset showedan increased F1-score up to 0.1 and classification accuracy upto 3 %. For 10 % training data, both metrics decreased. We present a deep convolutional GAN capable of creatingsynthetic data for the classification of craniosynostosis. Us-ing a moderate amount of synthetic data using a GAN showedslightly better performance, but had little effect overall. Thesimulated scarcity scenario of 10 % training data may havelimited the model’s ability to learn the underlying data distribution.

Abstract:

Objective: To investigate cardiac activation maps estimated using electrocardiographic imaging and to find methods reducing line-of-block (LoB) artifacts, while preserving real LoBs. Methods: Body surface potentials were computed for 137 simulated ventricular excitations. Subsequently, the inverse problem was solved to obtain extracellular potentials (EP) and transmembrane voltages (TMV). From these, activation times (AT) were estimated using four methods and compared to the ground truth. This process was evaluated with two cardiac mesh resolutions. Factors contributing to LoB artifacts were identified by analyzing the impact of spatial and temporal smoothing on the morphology of source signals. Results: AT estimation using a spatiotemporal derivative performed better than using a temporal derivative. Compared to deflection-based AT estimation, correlation-based methods were less prone to LoB artifacts but performed worse in identifying real LoBs. Temporal smoothing could eliminate artifacts for TMVs but not for EPs, which could be linked to their temporal morphology. TMVs led to more accurate ATs on the septum than EPs. Mesh resolution had a negligible effect on inverse reconstructions, but small distances were important for cross-correlation-based estimation of AT delays. Conclusion: LoB artifacts are mainly caused by the inherent spatial smoothing effect of the inverse reconstruction. Among the configurations evaluated, only deflection-based AT estimation in combination with TMVs and strong temporal smoothing can prevent LoB artifacts, while preserving real LoBs. Significance: Regions of slow conduction are of considerable clinical interest and LoB artifacts observed in non-invasive ATs can lead to misinterpretations. We addressed this problem by identifying factors causing such artifacts and methods to reduce them.

Abstract:

Left atrial enlargement (LAE) is one of the risk factors for atrial fibrillation (AF). A non-invasive and automated detection of LAE with the 12-lead electrocardiogram (ECG) could therefore contribute to an improved AF risk stratification and an early detection of new-onset AF incidents. However, one major challenge when applying machine learning techniques to identify and classify cardiac diseases usually lies in the lack of large, reliably labeled and balanced clinical datasets. We therefore examined if the extension of clinical training data by simulated ECGs derived from a novel bi-atrial shape model could improve the automated detection of LAE based on P waves of the 12-lead ECG. We derived 95 volumetric geometries from the bi-atrial statistical shape model with continuously increasing left atrial volumes in the range of 30 ml to 65 ml. Electrophysiological simulations with 10 different conduction velocity settings and 2 different torso models were conducted. Extracting the P waves of the 12-lead ECG thus yielded a synthetic dataset of 1,900 signals. Besides the simulated data, 7,168 healthy and 309 LAE ECGs from a public clinical ECG database were available for training and testing of an LSTM network to identify LAE. The class imbalance of the training data could be reduced from 1:23 to 1:6 when adding simulated data to the training set. The accuracy evaluated on the test dataset comprising a subset of the clinical ECG recordings improved from 0.91 to 0.95 if simulated ECGs were included as an additional input for the training of the classifier. Our results suggest that using a bi-atrial statistical shape model as a basis for ECG simulations can help to overcome the drawbacks of clinical ECG recordings and can thus lead to an improved performance of machine learning classifiers to detect LAE based on the 12-lead ECG.

Abstract:

Cranio-maxillofacial surgery often alters the aesthetics of the face which can be a heavy burden for patients to decide whether or not to undergo surgery. Today, physicians can predict the post-operative face using surgery planning tools to support the patient’s decision-making. While these planning tools allow a simulation of the post-operative face, the facial texture must usually be captured by another 3D texture scan and subsequently mapped on the simulated face. This approach often results in face predictions that do not appear realistic or lively looking and are therefore ill-suited to guide the patient’s decision-making. Instead, we propose a method using a generative adversarial network to modify a facial image according to a 3D soft-tissue estimation of the post-operative face. To circumvent the lack of available data pairs between pre- and post-operative measurements we propose a semi-supervised training strategy using cycle losses that only requires paired open-source data of images and 3D surfaces of the face’s shape. After training on “in-the-wild” images we show that our model can realistically manipulate local regions of a face in a 2D image based on a modified 3D shape. We then test our model on four clinical examples where we predict the post-operative face according to a 3D soft-tissue prediction of surgery outcome, which was simulated by a surgery planning tool. As a result, we aim to demonstrate the potential of our approach to predict realistic post-operative images of faces without the need of paired clinical data, physical models, or 3D texture scans.

Abstract:

Activation times (AT) describe the sequence of cardiac depolarization and represent one of the most important parameters for analysis of cardiac electrical activity. However, estimation of ATs can be challenging due to multiple sources of noise such as fractionation or baseline wander. If ATs are estimated from signals reconstructed using electrocardiographic imaging (ECGI), additional problems can arise from over-smoothing or due to ambiguities in the inverse problem. Often, resulting AT maps show falsely homogeneous regions or artificial lines of block. As ATs are not only important clinically, but are also commonly used for evaluation of ECGI methods, it is important to understand where these errors come from. We present results from a community effort to compare methods for AT estimation on a common dataset of simulated ventricular pacings. ECGI reconstructions were performed using three different surface source models: transmembrane voltages, epi-endo potentials and pericardial potentials, all using 2nd-order Tikhonov and 6 different regularization parameters. ATs were then estimated by the community participants and compared to the ground truth. While the pacing site had the largest effect on AT correlation coefficients (CC larger for lateral than for septal pacings), there were also differences between methods and source models that were poorly reflected in CCs. Results indicate that artificial lines of block are most severe for purely temporal methods. Compared to the other source models, ATs estimated from transmembrane voltages are more precise and less prone to artifacts.