Christoph Haarburger

Diffusion MRI Signal Augmentation: From Single Shell to Multi Shell with Deep Learning

High Angular Resolution Diffusion Imaging makes it possible to capture information about the course and location of complex fiber structures in the human brain. Ideally, multi-shell sampling would be applied, which however increases the acquisition time. Therefore, multi-shell acquisitions are considered infeasible for practical use in a clinical setting. In this work, we present a data-driven approach that is able to augment single-shell signals to multi-shell signals based on Deep Neural Networks and Spherical Harmonics. The proposed concept is evaluated on synthetic data to investigate the impact of noise and number of gradients. Moreover, it is evaluated on human brain data from the Human Connectome Project, comprising 100 scans from different subjects. The proposed approach makes it possible to drastically reduce the signal acquisition time and performs equally well on both synthetic as well as real human brain data.

Towards an algorithm for automatic accelerometer-based pulse presence detection during cardiopulmonary resuscitation

Manual palpation is still the gold standard for assessment of pulse presence during cardiopulmonary resuscitation (CPR) for professional rescuers. However, this method is unreliable, time-consuming and subjective. Therefore, reliable, quick and objectified assessment of pulse presence in cardiac arrest situations to assist professional rescuers is still an unmet need. Accelerometers may present a promising sensor modality as pulse palpation technology for which pulse detection at the carotid artery has been demonstrated to be feasible. This study extends previous work by presenting an algorithm for automatic, accelerometer-based pulse presence detection at the carotid site during CPR. We show that accelerometers might be helpful in automated detection of pulse presence during CPR.

Transfer Learning for Breast Cancer Malignancy Classification based on Dynamic Contrast-Enhanced MR Images

In clinical contexts with very limited annotated data, such as breast cancer diagnosis, training state-of-the art deep neural networks is not feasible. As a solution, we transfer parameters of networks pretrained on natural RGB images to malignancy classification of breast lesions in dynamic contrast-enhanced MR images. Since DCE-MR images comprise several contrasts and timepoints, a direct finetuning of pretrained networks expecting three input channels is not possible. Based on the hypothesis that a subset of the acquired image data is sufficient for a computer-aided diagnosis, we provide an experimental comparison of all possible subsets of MR image contrasts and determine the best combination for malignancy classification. A subset of images acquired at three timepoints of dynamic T1-weighted images which closely corresponds to human interpretation performs best with an AUC of 0.839.

Reliable Estimation of the Number of Compartments in Diffusion MRI

A-priori knowledge of the number of fibers in a voxel is mandatory and crucial when reconstructing multi-fiber voxels in diffusion MRI. Especially for clinical purposes, this estimation needs to be stable, even when only few gradient directions are acquired. In this work, we propose a novel approach to address this problem based on a deep convolutional neural network (CNN), which is able to identify important gradient directions and can be directly trained on real data. To obtain a ground truth using real data, 100 uncorrelated Human Connectome Project datasets are utilized, with a state-of-the-art framework used for generating a relative ground truth. It is shown that this CNN approach outperforms other state-of-the-art machine learning approaches.

Segmentierung von Zellkernen für Hochdurchsatz-DNA-Bildzytomerie

Krebs lasst sich durch das Vermessen des DNA-Gehaltes morphologisch auffalliger Zellkerne fruhzeitig diagnostizieren. Die manuelle Erfassung von Zellkernen ist sehr arbeitsaufwandig, lasst sich jedoch durch ein virtuelles Mikroskop und ein Mustererkennungssystem beschleunigen. Die Umstellung auf die Hochdurchsatz-Variante erfordert eine neuartige Zellkern-Segmentierung, die effizient und prazise sein muss und moglichst wenige irrelevante Objekte segmentiert. Zu diesem Zweck wurde ein dreistufiges Verfahren entwickelt: Pixel werden anhand ihrer Farbwerte mittels eines Maximum-Likelihood-Ansatzes oder durch einen kNN, SVM oder Adaboost-Klassifikator klassifiziert und bilden eine initiale Segmentierung. Irrelevante Objekte werden anhand schnell zu berechnender Merkmale ausgeschlossen. Ein weiteres Merkmal wird berechnet, um zu entscheiden, ob die Kontur einer Verbesserung bedarf. Falls erforderlich geschieht dies durch ein parametrisches Aktives Konturmodell. Auf einem Testset von 80 annotierten Sichtfeldern erreicht die Segmentierung mittels kNN-Klassifikator die beste Performance. Zellkerne werden mit einer Sensitivitat von 98.9% detektiert. Im Vergleich zum bisherigen Vorgehen werden bis zu 33% weniger Objekte segmentiert, die keine Zellkerne sind. Nach der Verbesserung wird ein Dice-Koeffizient von 0,908 und eine Hausdorff-Distanz von 0,721 μm erreicht.