@article{AmiranashviliLuedkeLietal.2024, author = {Amiranashvili, Tamaz and L{\"u}dke, David and Li, Hongwei Bran and Zachow, Stefan and Menze, Bjoern}, title = {Learning continuous shape priors from sparse data with neural implicit functions}, volume = {94}, journal = {Medical Image Analysis}, doi = {10.1016/j.media.2024.103099}, pages = {103099}, year = {2024}, abstract = {Statistical shape models are an essential tool for various tasks in medical image analysis, including shape generation, reconstruction and classification. Shape models are learned from a population of example shapes, which are typically obtained through segmentation of volumetric medical images. In clinical practice, highly anisotropic volumetric scans with large slice distances are prevalent, e.g., to reduce radiation exposure in CT or image acquisition time in MR imaging. For existing shape modeling approaches, the resolution of the emerging model is limited to the resolution of the training shapes. Therefore, any missing information between slices prohibits existing methods from learning a high-resolution shape prior. We propose a novel shape modeling approach that can be trained on sparse, binary segmentation masks with large slice distances. This is achieved through employing continuous shape representations based on neural implicit functions. After training, our model can reconstruct shapes from various sparse inputs at high target resolutions beyond the resolution of individual training examples. We successfully reconstruct high-resolution shapes from as few as three orthogonal slices. Furthermore, our shape model allows us to embed various sparse segmentation masks into a common, low-dimensional latent space — independent of the acquisition direction, resolution, spacing, and field of view. We show that the emerging latent representation discriminates between healthy and pathological shapes, even when provided with sparse segmentation masks. Lastly, we qualitatively demonstrate that the emerging latent space is smooth and captures characteristic modes of shape variation. We evaluate our shape model on two anatomical structures: the lumbar vertebra and the distal femur, both from publicly available datasets.}, language = {en} } @misc{Peter2023, type = {Master Thesis}, author = {Peter, Clea}, title = {Improving the Realism of Synthetic Cryogenic Electron Micrographs Using Generative Adversarial Networks}, year = {2023}, abstract = {This thesis addresses the problem of synthetic-to-real image refinement applied to tilt series of cryogenic electron micrographs. It explores the possibility of improving the realism of synthesized micrographs using generative adversarial networks, which could help to improve the automatic segmentation of cellular structures based on deep learning methods. For image refinement, three image-to-image translation networks were used to transfer the appearance of real micrographs to synthetic micrographs while preserving their original content, including the location and shape of particles. The first model, called SimGAN, was unable to produce any meaningful refinement. Instead, the content of the synthetic micrographs was corrupted by the addition of extensive noise, making SimGAN unsuitable for the problem of this thesis. As a result, CycleGAN was introduced and its refinement of synthetic micrographs matches the appearance of real micrographs very well. However, structural changes in the position and shape of particles were observed after translation. To avoid this behavior, CUT was used as a third model on an exploratory basis but its performance was inferior to that of CycleGAN. In conclusion, CycleGAN proved to be the most promising image-to-image translation model for the images presented, although it does not solve the main problem of this thesis. In order to do so, further modifications, such as the addition of a structural constraint during translation, are required.}, language = {en} } @phdthesis{Tack2024, author = {Tack, Alexander}, title = {Machine Learning-based Assessment of Multiple Anatomical Structures in Medical Image Data for Diagnosis and Prediction of Knee Osteoarthritis}, doi = {10.14279/depositonce-19738}, year = {2024}, abstract = {Knee osteoarthritis (KOA) is a degenerative disease that leads to pain and loss of function. It is estimated to affect over 500 million humans world-wide and is one of the most common reasons for disability. KOA is usually diagnosed by radiologists or clinical experts by anamnesis, physical examination, and by assessing medical image data. The latter is typically acquired using X-Ray or magnetic resonance imaging. Since manual image reading is subjective, tedious and time-consuming, automated methods are required for a fast and objective decision support and for a better understanding of the pathogenesis of KOA. This thesis sets a foundation towards automated computation of image-based KOA biomarkers for holistic assessment of the knee. This involves the assessment of multiple knee bones and soft tissues. An assessment of particular structures requires localization of these tissues. In order to automate a faithful localization of anatomical structures, deep learning-based methods are investigated and utilized. Additionally, convolutional neural networks (CNNs) are used for classification of medical image data, i.e., for a direct determination of the disease status and to detect anatomical structures and landmarks. The automatically computed anatomical volumes, locations, and other measurements are finally compared to values acquired by clinical experts and evaluated for clustering of KOA groups, classification of KOA severity, prediction of KOA progression, and prediction of total knee replacement. In various experiments it is shown that CNN-based methods are suitable for accurate medical image segmentation, object detection, landmark detection, and direct classification of disease stages from the image data. Computed features related to the menisci are found to be most expressive in terms of clustering of KOA groups and predicting of future disease states, thus allowing diagnosis of current KOA conditions and prediction of future conditions. The conclusion of this thesis is that machine learning-based, fully automated processing of medical image data shows potential for diagnosis and prediction of KOA grades. Future studies could investigate additional features in order to achieve an assessment of the whole knee or validate the findings of this work in clinical studies.}, language = {en} } @inproceedings{AmiranashviliLuedkeLietal.2022, author = {Amiranashvili, Tamaz and L{\"u}dke, David and Li, Hongwei and Menze, Bjoern and Zachow, Stefan}, title = {Learning Shape Reconstruction from Sparse Measurements with Neural Implicit Functions}, booktitle = {Medical Imaging with Deep Learning}, year = {2022}, abstract = {Reconstructing anatomical shapes from sparse or partial measurements relies on prior knowledge of shape variations that occur within a given population. Such shape priors are learned from example shapes, obtained by segmenting volumetric medical images. For existing models, the resolution of a learned shape prior is limited to the resolution of the training data. However, in clinical practice, volumetric images are often acquired with highly anisotropic voxel sizes, e.g. to reduce image acquisition time in MRI or radiation exposure in CT imaging. The missing shape information between the slices prohibits existing methods to learn a high-resolution shape prior. We introduce a method for high-resolution shape reconstruction from sparse measurements without relying on high-resolution ground truth for training. Our method is based on neural implicit shape representations and learns a continuous shape prior only from highly anisotropic segmentations. Furthermore, it is able to learn from shapes with a varying field of view and can reconstruct from various sparse input configurations. We demonstrate its effectiveness on two anatomical structures: vertebra and femur, and successfully reconstruct high-resolution shapes from sparse segmentations, using as few as three orthogonal slices.}, language = {en} } @article{ObermeierHeimBiereetal.2022, author = {Obermeier, Patrick E and Heim, Albert and Biere, Barbara and Hage, Elias and Alchikh, Maren and Conrad, Tim and Schweiger, Brunhilde and Rath, Barbara A}, title = {Linking digital surveillance and in-depth virology to study clinical patterns of viral respiratory infections in vulnerable patient populations}, volume = {25}, journal = {iScience}, number = {5}, publisher = {Cell Press}, doi = {10.1016/j.isci.2022.104276}, year = {2022}, abstract = {To improve the identification and management of viral respiratory infections, we established a clinical and virologic surveillance program for pediatric patients fulfilling pre-defined case criteria of influenza-like illness and viral respiratory infections. The program resulted in a cohort comprising 6,073 patients (56\% male, median age 1.6 years, range 0-18.8 years), where every patient was assessed with a validated disease severity score at the point-of-care using the ViVI ScoreApp. We used machine learning and agnostic feature selection to identify characteristic clinical patterns. We tested all patients for human adenoviruses, 571 (9\%) were positive. Adenovirus infections were particularly common and mild in children ≥1 month of age but rare and potentially severe in neonates: with lower airway involvement, disseminated disease, and a 50\% mortality rate (n = 2/4). In one fatal case, we discovered a novel virus …}, language = {en} } @article{HajarolasvadiSunkaraKhavnekaretal.2022, author = {Hajarolasvadi, Noushin and Sunkara, Vikram and Khavnekar, Sagar and Beck, Florian and Brandt, Robert and Baum, Daniel}, title = {Volumetric macromolecule identification in cryo-electron tomograms using capsule networks}, volume = {23}, journal = {BMC Bioinformatics}, number = {360}, doi = {10.1186/s12859-022-04901-w}, year = {2022}, abstract = {Background: Despite recent advances in cellular cryo-electron tomography (CET), developing automated tools for macromolecule identification in submolecular resolution remains challenging due to the lack of annotated data and high structural complexities. To date, the extent of the deep learning methods constructed for this problem is limited to conventional Convolutional Neural Networks (CNNs). Identifying macromolecules of different types and sizes is a tedious and time-consuming task. In this paper, we employ a capsule-based architecture to automate the task of macro- molecule identification, that we refer to as 3D-UCaps. In particular, the architecture is composed of three components: feature extractor, capsule encoder, and CNN decoder. The feature extractor converts voxel intensities of input sub-tomograms to activities of local features. The encoder is a 3D Capsule Network (CapsNet) that takes local features to generate a low-dimensional representation of the input. Then, a 3D CNN decoder reconstructs the sub-tomograms from the given representation by upsampling. Results: We performed binary and multi-class localization and identification tasks on synthetic and experimental data. We observed that the 3D-UNet and the 3D-UCaps had an F1-score mostly above 60\% and 70\%, respectively, on the test data. In both network architectures, we observed degradation of at least 40\% in the F1-score when identifying very small particles (PDB entry 3GL1) compared to a large particle (PDB entry 4D8Q). In the multi-class identification task of experimental data, 3D-UCaps had an F1-score of 91\% on the test data in contrast to 64\% of the 3D-UNet. The better F1-score of 3D-UCaps compared to 3D-UNet is obtained by a higher precision score. We speculate this to be due to the capsule network employed in the encoder. To study the effect of the CapsNet-based encoder architecture further, we performed an ablation study and perceived that the F1-score is boosted as network depth is increased which is in contrast to the previously reported results for the 3D-UNet. To present a reproducible work, source code, trained models, data as well as visualization results are made publicly available. Conclusion: Quantitative and qualitative results show that 3D-UCaps successfully perform various downstream tasks including identification and localization of macro- molecules and can at least compete with CNN architectures for this task. Given that the capsule layers extract both the existence probability and the orientation of the molecules, this architecture has the potential to lead to representations of the data that are better interpretable than those of 3D-UNet.}, language = {en} } @inproceedings{LuedkeAmiranashviliAmbellanetal.2022, author = {L{\"u}dke, David and Amiranashvili, Tamaz and Ambellan, Felix and Ezhov, Ivan and Menze, Bjoern and Zachow, Stefan}, title = {Landmark-free Statistical Shape Modeling via Neural Flow Deformations}, volume = {13432}, booktitle = {Medical Image Computing and Computer Assisted Intervention - MICCAI 2022}, publisher = {Springer, Cham}, arxiv = {http://arxiv.org/abs/2209.06861}, doi = {10.1007/978-3-031-16434-7_44}, year = {2022}, abstract = {Statistical shape modeling aims at capturing shape variations of an anatomical structure that occur within a given population. Shape models are employed in many tasks, such as shape reconstruction and image segmentation, but also shape generation and classification. Existing shape priors either require dense correspondence between training examples or lack robustness and topological guarantees. We present FlowSSM, a novel shape modeling approach that learns shape variability without requiring dense correspondence between training instances. It relies on a hierarchy of continuous deformation flows, which are parametrized by a neural network. Our model outperforms state-of-the-art methods in providing an expressive and robust shape prior for distal femur and liver. We show that the emerging latent representation is discriminative by separating healthy from pathological shapes. Ultimately, we demonstrate its effectiveness on two shape reconstruction tasks from partial data. Our source code is publicly available (https://github.com/davecasp/flowssm).}, language = {en} } @article{LelievreZhang2019, author = {Leli{\`e}vre, Tony and Zhang, Wei}, title = {Pathwise estimates for effective dynamics: the case of nonlinear vectorial reaction coordinates}, journal = {Multiscale Modeling and Simulation}, number = {17}, arxiv = {http://arxiv.org/abs/1805.01928}, doi = {10.1137/18M1186034}, pages = {1019 -- 1051}, year = {2019}, abstract = {Effective dynamics using conditional expectation was proposed in [F. Legoll and T. Leli{\`e}vre, Nonlinearity, 2010] to approximate the essential dynamics of high-dimensional diffusion processes along a given reaction coordinate. The approximation error of the effective dynamics when it is used to approximate the behavior of the original dynamics has been considered in recent years. As a continuation of the previous work [F. Legoll, T. Leli{\`e}vre, and S. Olla, Stoch. Process. Appl, 2017], in this paper we obtain pathwise estimates for effective dynamics when the reaction coordinate function is either nonlinear or vector-valued.}, language = {en} } @article{HartmannSchuetteZhang2019, author = {Hartmann, Carsten and Sch{\"u}tte, Christof and Zhang, Wei}, title = {Jarzynski's equality, fluctuation theorems, and variance reduction: Mathematical analysis and numerical algorithms}, volume = {175}, journal = {Journal of Statistical Physics}, number = {6}, arxiv = {http://arxiv.org/abs/1803.09347}, doi = {10.1007/s10955-019-02286-4}, pages = {1214 -- 1261}, year = {2019}, abstract = {In this paper, we study Jarzynski's equality and fluctuation theorems for diffusion processes. While some of the results considered in the current work are known in the (mainly physics) literature, we review and generalize these nonequilibrium theorems using mathematical arguments, therefore enabling further investigations in the mathematical community. On the numerical side, variance reduction approaches such as importance sampling method are studied in order to compute free energy differences based on Jarzynski's equality.}, language = {en} } @article{ZhangKlusConradetal.2019, author = {Zhang, Wei and Klus, Stefan and Conrad, Tim and Sch{\"u}tte, Christof}, title = {Learning chemical reaction networks from trajectory data}, volume = {18}, journal = {SIAM Journal on Applied Dynamical Systems (SIADS)}, number = {4}, arxiv = {http://arxiv.org/abs/1902.04920}, doi = {10.1137/19M1265880}, pages = {2000 -- 2046}, year = {2019}, abstract = {We develop a data-driven method to learn chemical reaction networks from trajectory data. Modeling the reaction system as a continuous-time Markov chain and assuming the system is fully observed,our method learns the propensity functions of the system with predetermined basis functions by maximizing the likelihood function of the trajectory data under l^1 sparse regularization. We demonstrate our method with numerical examples using synthetic data and carry out an asymptotic analysis of the proposed learning procedure in the infinite-data limit.}, language = {en} }