Publikationen

Frequent complications arising from low anterior resections include urinary and fecal incontinence, as well as sexual disorders, which are commonly associated with damage to the pelvic autonomic nerves during surgery. To assist the surgeon in preserving pelvic autonomic nerves, a novel approach for intraoperative pelvic neuromonitoring was investigated that is based on impedance measurements of the innervated organs. The objective of this work was to develop an algorithm called AMINA to classify the bioimpedance signals, with the goal of facilitating signal interpretation for the surgeon. Thirty patients included in a clinical investigation underwent nerve-preserving robotic rectal surgery using intraoperative pelvic neuromonitoring. Contraction of the urinary bladder and/or rectum, triggered by direct stimulation of the innervating nerves, resulted in a change in tissue impedance signal, allowing the nerves to be identified and preserved. Impedance signal characteristics in the time domain and the time-frequency domain were calculated and classified to develop the AMINA. Stimulation-induced positive impedance changes were statistically significantly different from negative stimulation responses by the percent amplitude of impedance change Amax in the time domain. Positive impedance changes and artifacts were distinguished by classifying wavelet scales resulting from peak detection in the continuous wavelet transform scalogram, which allowed implementation of a decision tree underlying the AMINA. The sensitivity of the software-based signal evaluation by the AMINA was 96.3%, whereas its specificity was 91.2%. This approach streamlines and automates the interpretation of impedance signals during intraoperative pelvic neuromonitoring.

There is evidence that humans can survive a direct lightning strike to the head. Our question is: could water (rain) on the skin contribute to an increase in the survival rate? We measure the influence of rain during high-energy direct lightning strikes on a realistic three-compartment human head phantom. We find a lower number of perforations and eroded areas near the lightning strike impact points on the head phantom when rain was applied compared to no rain. Current amplitudes in the brain were lower with rain compared to no rain before a fully formed flashover. We conclude that rain on the scalp potentially contributes to the survival rate of 70-90% due to: (1) lower current exposition in the brain before a fully formed flashover, and (2) reduced mechanical and thermal damage.

Today, the process of producing a printed circuit board assembly (PCBA) is growing rapidly, and this process requires cutting-edge debugging and testing of the boards. The Automatic Optical Inspection (AOI) process detects defects in the boards, components, or solder pads using image processing and machine learning (ML) algorithms. Although state-of-the-art approaches for identifying defects are well developed, due to three main issues, the ML algorithms and datasets are incapable of fully integrating into industrial plants. These issues are privacy limitations for sharing data, the distribution shifts in the PCBA industry, and the absence of a degree of freedom for reproducible and modifiable synthetic datasets.

Background: One of the limitations in leveraging the potential of artificial intelligence in X-ray imaging is the limited availability of annotated training data. As X-ray and CT shares similar imaging physics, one could achieve cross-domain data sharing, so to generate labeled synthetic X-ray images from annotated CT volumes as digitally reconstructed radiographs (DRRs). To account for the lower resolution of CT and the CT-generated DRRs as compared to the real X-ray images, we propose the use of super-resolution (SR) techniques to enhance the CT resolution before DRR generation. Purpose: As spatial resolution can be defined by the modulation transfer function kernel in CT physics, we propose to train a SR network using paired low-resolution (LR) and high-resolution (HR) images by varying the kernel's shape and cutoff frequency. This is different to previous deep learning-based SR techniques on RGB and medical images which focused on refining the sampling grid. Instead of generating LR images by bicubic interpolation, we aim to create realistic multi-detector CT (MDCT) like LR images from HR cone-beam CT (CBCT) scans. Methods: We propose and evaluate the use of a SR U-Net for the mapping between LR and HR CBCT image slices. We reconstructed paired LR and HR training volumes from the same CT scans with small in-plane sampling grid size of 0.20 x 0.20 mm2. We used the residual U-Net architecture to train two models. SRUN K Res: trained with kernel-based LR images, and SRUN I Res: trained with bicubic downsampled data as baseline. Both models are trained on one CBCT dataset (n = 13 391). The performance of both models was then evaluated on unseen kernel-based and interpolation-based LR CBCT images (n = 10 950), and also on MDCT images (n = 1392). Results: Five-fold cross validation and ablation study were performed to find the optimal hyperparameters. Both SRUN K Res and SRUN I Res models show significant improvements (p-value < 0.05) in mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and structural similarity index measures (SSIMs) on unseen CBCT images. Also, the improvement percentages in MAE, PSNR, and SSIM by SRUN K Res is larger than SRUN I Res. For SRUN K Res, MAE is reduced by 14%, and PSNR and SSIMs increased by 6 and 8%, respectively. To conclude, SRUN K Res outperforms SRUN I Res, which the former generates sharper images when tested with kernel-based LR CBCT images as well as cross-modality LR MDCT data. Conclusions: Our proposed method showed better performance than the baseline interpolation approach on unseen LR CBCT. We showed that the frequency behavior of the used data is important for learning the SR features. Additionally, we showed cross-modality resolution improvements to LR MDCT images. Our approach is, therefore, a first and essential step in enabling realistic high spatial resolution CT-generated DRRs for deep learning training.

Software analytics integrated with complex databases can deliver project intelligence into the hands of software engineering (SE) experts for satisfying their information needs. A new and promising machine learning technique known as text-to-SQL automatically extracts information for users of complex databases without the need to fully understand the database structure nor the accompanying query language. Users pose their request as so-called natural language utterance, i.e., question. Our goal was evaluating the performance and applicability of text-to-SQL approaches on data derived from tools typically used in the workflow of software engineers for satisfying their information needs. We carefully selected and discussed five seminal as well as state-of-the-art text-to-SQL approaches and conducted a comparative assessment using the large-scale, cross-domain Spider dataset and the SE domain-specific SEOSS-Queries dataset. Furthermore, we study via a survey how SE professionals perform in satisfying their information needs and how they perceive text-to-SQL approaches. For the best performing approach, we observe a high accuracy of 94% in query prediction when training specifically on SE data. This accuracy is almost independent of the query’s complexity. At the same time, we observe that SE professionals have substantial deficits in satisfying their information needs directly via SQL queries. Furthermore, SE professionals are open for utilizing text-to-SQL approaches in their daily work, considering them less time-consuming and helpful. We conclude that state-of-the-art text-to-SQL approaches are applicable in SE practice for day-to-day information needs.