Optical coherence tomography angiography (OCTA) is a non-invasive imaging modality that can acquire high-resolution volumes of the retinal vasculature and aid the diagnosis of ocular, neurological and cardiac diseases. Segmenting the visible blood vessels is a common first step when extracting quantitative biomarkers from these images. Classical segmentation algorithms based on thresholding are strongly affected by image artifacts and limited signal-to-noise ratio. The use of modern, deep learning-based segmentation methods has been inhibited by a lack of large datasets with detailed annotations of the blood vessels. To address this issue, recent work has employed transfer learning, where a segmentation network is trained on synthetic OCTA images and is then applied to real data. However, the previously proposed simulations fail to faithfully model the retinal vasculature and do not provide effective domain adaptation. Because of this, current methods are unable to fully segment the retinal vasculature, in particular the smallest capillaries. In this work, we present a lightweight simulation of the retinal vascular network based on space colonization for faster and more realistic OCTA synthesis. We then introduce three contrast adaptation pipelines to decrease the domain gap between real and artificial images. We demonstrate the superior segmentation performance of our approach in extensive quantitative and qualitative experiments on three public datasets that compare our method to traditional computer vision algorithms and supervised training using human annotations. Finally, we make our entire pipeline publicly available, including the source code, pretrained models, and a large dataset of synthetic OCTA images.
9 February 2023 in Science of The Total Environment
Anaerobic digestion is a well-established tool at wastewater treatment plants for processing raw sludge; it can also be used to generate renewable energy by harvesting biogas in anaerobic digesters. Operational parameters, such as temperature, are usually set by plant operators according to expert knowledge. To completely utilize the potential of operational management, in this study, we calibrated a novel Temporal Fusion Transformer based on six years of life-scale time series data together with categorical features such as public holidays. The model design allows for the interpretability of the output in contrast to traditional data-driven techniques, using multi-head attention. In addition to forecasting the median biogas production rates for the following seven days, our model also yields quantiles, making it less prone to strong fluctuations. We used three well-known statistical techniques as benchmarks. The mean absolute percentage error of our forecasting approach is below 8 %.
1 October 2010 in ACM Transactions on Applied Perception
In this article, the data-driven haptic rendering approach presented in our earlier work is assessed. The approach relies on recordings from real objects from which a data-driven model is derived that captures the haptic properties of the object. We conducted two studies. In the first study, the Just Noticeable Difference (JND) for small forces, as encountered in our set-up, was determined. JNDs were obtained both for active and passive user interaction. A conservative threshold curve was derived that was then used to guide the model generation in the second study. The second study examined the achievable rendering fidelity for two objects with different stiffnesses. Subjects directly compared data-driven virtual feedback with the real objects. Results indicated that it is crucial to include dynamic material effects to achieve haptic feedback that cannot be distinguished from real objects. Results also showed that the fidelity is considerably decreased for stiffer objects due to limits of the display hardware.
We currently explore the application of haptic augmentation in the context of palpation training systems. The key idea is to modify real touch sensations with computed haptic feedback. In earlier work, we have introduced an algorithmic framework for determining appropriate augmentation forces during interaction at one contact point. In this paper, we present an extension of the approach to deal with manipulations at more than one contact location. At the heart of our method is the data-driven estimation of Hunt-Crossley model parameters in a pre-computation step. Feeding the parameters into a contact dynamics model allows us to approximate the feedback behavior of various physical tissue mock-ups. Further, we combine the parameter estimation with the tracking of the position of a stiffer inclusion in the mock-up. These data are employed to create a model of movement due to external forces. The combination of these models then allows us to represent and render the mutual effects at multiple contact points. Several experiments have been carried out on a setup with two haptic devices. Comparisons of recorded with simulated interaction data demonstrate the performance and potential of our method.
Haptic feedback in virtual reality-based trainers for surgical bone drilling is mostly provided via impedance-controlled haptic devices with limited displayable force. In addition, vibration feedback is often only of reduced fidelity. To overcome these shortcomings, we have developed a hand-held, actuated admittance-controlled drill mockup, comprising enhanced kinesthetic and tactile feedback. This paper reports on design and characterization of the device, and highlights its use for training. Kinesthetic feedback is provided through haptic augmentation, employing a ball-screw mechanism acting on a retractable drill-bit. Feedback computation relies on admittance control allowing for stable display of very high resistance forces, and thus material stiffness, which cannot be achieved with standard impedance-control approaches. For the tactile mechanism, a modified linear vibration actuator is directly attached to the mockup handle, improving signal transmission. To compensate the actuator dynamic, frequency-specific gains are adjusted in real-time. The performance of the device is characterized in several experiments, including comparisons to drilling with a real drill into artificial bone samples. We illustrate the capability of the mockup to render bone samples with different material layer stiffness and thickness. Moreover, we show that the mockup system allows for the same training effect as when rehearsing with a real drill.
In this work we extend ideas of machine learning to the domain of data-driven haptic rendering. The proposed approach facilitates the processing of high-dimensional haptic interaction signals, which so far proved too difficult for existing data-driven methods. The key idea is to construct a compact feature space in the frequency domain which allows for efficient data reduction via a feature selection process. First, in a recording stage, extensive force and displacement datasets are acquired in automated measurements on deformable sample objects. These data are then transformed into a dimensionally reduced, compact frequency space representation. Next, feature-based learning is carried out in this feature space to significantly reduce the size of the original dataset. Based on this, time-domain haptic models capable of real-time performance are finally generated to encode the forces arising from bimanual object interactions. The presented processing chain is generally applicable and extendable to more complex interactions with even higher-dimensional data. The resulting haptic models are directly usable for data-driven haptic rendering. We illustrate the improved performance in comparison with previously existing data-processing approaches.