Mirco Fuchs

Despite achieving high volumetric overlap, pixel-based deep learning segmentation approaches do not ensure true vascular connectivity, leading to fragmentations structurally analogous to Branch Retinal Artery Occlusions. This fundamental deficiency means that even segmentations achieving high pixel-wise overlap metrics (such as the Dice Score) do not guarantee or improve true vascular connectivity, rendering them unreliable for quantitative analysis and subsequent clinical risk stratification of high-risk systemic events, such as stroke. To resolve this, we introduce the Differentiable Endpoint Connectivity Loss (DECL), a novel, fully differentiable loss function that directly optimizes vascular continuity by precisely targeting errors in vessel endpoints. DECL leverages a differentiable soft-skeletonization module to impose two specialized penalties: a Normalized Count Penalty (L_c), which rigorously penalizes the discrete difference in the number of predicted versus true endpoints (ΔN_p), and a Normalized Distance Loss (L_d), which penalizes the Euclidean distance between the weighted centers of the true and predicted endpoints. In comparative experiments across four datasets (DRIVE, CHASE DB1, STARE, and OCTA-500), we tested DECL against established methods including Dice CE, clDice, cbDice, and SAC Loss. DECL and its hybrid variant, DECL_clDice, consistently recorded the highest mean Dice Score and clDice scores, specifically, DECL_clDiced improved the mean clDice score from 0.839 to 0.851 on CHASE DB1 and from 0.836 to 0.842 on DRIVE, while minimizing Betti Number Error (β-Error) andVariation of Information. The DECL method establishes a robust, mathematically justified framework for generating high-fidelity vascular maps, essential for accurate biomarker quantification and reliable clinical assessment.

Goal: We introduce a continuous, multimodal pain classification technique that utilizes camera-based data conducted in clinical settings. Methods: We integrate facial Action Units (AUs) obtained from samples with sequential vital parameters extracted from video data, and systematically validate the practicality of measuring heart rate variability (HRV) from video-derived photoplethysmographic signals against traditional sensor-based electrocardiogram measurements. Video-based AUs and HRV metrics acquired from ultra-short-term processing are combined into an automated, contactless, multimodal algorithm for binary pain classification. Utilizing logistic regression alongside leave-one-out cross-validation, this approach is developed and validated using the BioVid Heat Pain Database and subsequently tested with our surgical Individual Patient Data. Results: We achieve an F1-score of 53% on the BioVid Heat Pain Database and 48% on our Individual Patient Data with ultra-short-term processing. Conclusion: Our approach provides a robust foundation for future multimodal pain classification utilizing vital signs and mimic parameters from 5.5 s camera recordings.

Accurate spectral estimation from irregularly sampled signals is a long-standing challenge in signal processing, particularly in biomedical applications. Traditional methods, such as the Fast Fourier Transform (FFT), rely on resampling, which introduces artefacts and computational inefficiencies. To address this, we propose DySpect-Net (Dynamic Spectrum Network), a neural network-based model that directly predicts spectral representations from non-uniformly sampled data, eliminating the need for resampling. Our approach is validated using synthetic datasets that mimic real-world sampling irregularities. Unlike conventional spectral estimation methods, DySpect-Net learns to recognise and reconstruct dominant frequency components dynamically. We introduce FAVE (Frequency Absolute Variance Error), a novel evaluation metric designed to quantify spectral peak alignment accuracy, outperforming conventional metrics such as Mean Squared Error (MSE). Comparative analysis demonstrates that DySpect-Net performs superior spectral estimation, particularly in low-sampling-rate conditions, with fewer distortions than FFT-based approaches. The proposed model has broad applications in biomedical signal analysis, e.g. remote photoplethysmography (rPPG), and other domains where sampling irregularities compromise spectral fidelity. Future research will focus on integrating FAVE as a loss function to enhance spectral reconstruction accuracy.

This work concerns itself with the task of reconstructing all edges of an arbitrary 3D wire-frame model projected to an image plane. We explore a bottom-up part-wise procedure undertaken by an RL agent to segment and reconstruct these 2D multipart objects. The environment's state is represented as a four-colour image, where different colours correspond to background, a target edge, a reconstruction line, and the overlap of both. At each step, the agent can transform the reconstruction line within a four-dimensional action space or terminate the episode using a specific termination action. To investigate the impact of reward function formulations, we tested episodic and incremental rewards, as well as combined approaches. Empirical results demonstrated that the latter yielded the most effective training performance. To further enhance efficiency and stability, we introduce curriculum learning strategies. First, an action-based curriculum was implemented, where the agent was initially restricted to a reduced action space, being able to only perform three of the five possible actions, before progressing to the full action space. Second, we test a task-based curriculum, where the agent first solves a simplified version of the problem before being presented with the full, more complex task. This second approach produced promising results, as the agent not only successfully transitioned from learning the simplified task to mastering the full task, but in doing so gained significant performance. This study demonstrates the potential of an iterative RL wire-frame reconstruction in two dimensions. By combining optimized reward function formulations with curriculum learning strategies, we achieved significant improvements in training success. The proposed methodology provides an effective framework for solving similar tasks and represents a promising direction for future research in the field.

Additive manufacturing is gaining importance in a number of application areas, and there is an increased demand for mechanically resilient components. A way to improve the mechanical properties of parts made of thermoplastics is by using reinforcing material. The study demonstrates the development of a monitoring procedure for a fused filament fabrication-based co-extrusion process for manufacturing wire-reinforced thermoplastic components. Test components in two variants are produced, and data acquisition is carried out with a laser line scanner. The collected data are employed to train deep neural networks to classify the printed layers, aiming for the deep neural networks to be able to classify four different classes and identify layers with insufficient quality. A dedicated convolutional neural network is designed taking into account various factors such as layer architecture, data pre-processing and optimization methods. Several network architectures, including transfer learning (based on VGG16 and ResNet50), with and without fine-tuning, are compared in terms of their performance based on the F1 score. Both the transfer learning model with ResNet50 and the fine-tuning model achieve an F1 score of 84% and 83%, respectively, for the decisive class ‘wire bad’ classifying inadequate reinforcement.

Exploiting photoplethysmography signals (PPG) for non-invasive blood pressure (BP) measurement is interesting for various reasons. First, PPG can easily be measured using fingerclip sensors. Second, camera based approaches allow to derive remote PPG (rPPG) signals similar to PPG and therefore provide the opportunity for non-invasive measurements of BP. Various methods relying on machine learning techniques have recently been published. Performances are often reported as the mean average error (MAE) on the data which is problematic. This work aims to analyze the PPG- and rPPG based BP prediction error with respect to the underlying data distribution. First, we train established neural network (NN) architectures and derive an appropriate parameterization of input segments drawn from continuous PPG signals. Second, we use this parameterization to train NNs with a larger PPG dataset and carry out a systematic evaluation of the predicted blood pressure. The analysis revealed a strong systematic increase of the prediction error towards less frequent BP values across NN architectures. Moreover, we tested different train/test set split configurations which underpin the importance of a careful subject-aware dataset assignment to prevent overly optimistic results. Third, we use transfer learning to train the NNs for rPPG based BP prediction. The resulting performances are similar to the PPG-only case. Finally, we apply different personalization techniques and retrain our NNs with subject-specific data for both the PPG-only and rPPG case. Whilst the particular technique is less important, personalization reduces the prediction errors significantly.

In this study we compared various neural network types for the task of automatic infant vocalization classification, i.e convolutional, recurrent and fully-connected networks as well as combinations of thereof. The goal was to first determine the optimal configuration for each network type to then identify the type with the highest overall performance. This investigation helps to employ neural networks more effectively to infant vocalization classification tasks, which typically offer low amounts of training data. To this end, we defined a unified neural network architecture scheme for audio classification from which we derived various network types. For each type we performed a semi-random hyperparameter search which employed regression trees to both focus the search space as well as derive insights on the most influential parameters. We finally compared the test performances of the best performing configurations in an contest-like setup. Our key findings are: (1) Networks with convolutional stages reached the highest performance, regardless of being combined with fully-connected or recurrent layers. (2) The most influential architectural hyperparameter for all types were the integration operations for reducing tensor dimensionality between network stages. The best performing configurations reached test performances of 75% unweighted average recall, surpassing previously published benchmarks.