The electrical activity of the diaphragm (EAdi) is a novel monitoring parameter for patients under assisted ventilation and is used for assessing the patient’s neural respiratory drive. It is recorded by an array of electrodes placed inside the esophagus at the level of the diaphragm. A noninvasive alternative is the measurement of the electromyogram by means of skin surface electrodes (sEMG). The respiratory sEMG signal, however, is subject to electrocardiographic interference and crosstalk from other muscles and may also pick up a different part of the muscular activity. In this work, we propose to use a deep neural network to predict the electrical activity of the diaphragm as measured by a nasogastric catheter from sEMG measurements. We use a ResNet based architecture and train the network to directly regress the EAdi as a supervised learning task - we further investigate a heatmap based regression approach. The proposed methods are evaluated on a clinical dataset consisting of 77 recordings from mechanically ventilated patients. For the direct regression task, the network’s predictions reach a Pearson correlation coefficient (PCC) of 0.818 with EAdi on the hold-out set. The heatmap regression increases the PCC to 0.830 while at the same time achieving a lower mean absolute error, indicating a superior performance. From our results we conclude that sEMG measurements may be used to predict the internal activity of the diaphragm as measured invasively using a nasogastric catheter.

Purpose During brain tumor surgery, care must be taken to accurately differentiate between tumorous and healthy tissue, as inadvertent resection of functional brain areas can cause severe consequences. Since visual assessment can be difficult during tissue resection, neurosurgeons have to rely on the mechanical perception of tissue, which in itself is inherently challenging. A commonly used instrument for tumor resection is the ultrasonic aspirator, whose system behavior is already dependent on tissue properties. Using data recorded during tissue fragmentation, machine learning-based tissue differentiation is investigated for the first time utilizing ultrasonic aspirators. Methods Artificial tissue model with two different mechanical properties is synthesized to represent healthy and tumorous tissue. 40,000 temporal measurement points of electrical data are recorded in a laboratory environment using a CNC machine. Three different machine learning approaches are applied: a random forest (RF), a fully connected neural network (NN) and a 1D convolutional neural network (CNN). Additionally, different preprocessing steps are investigated. Results Fivefold cross-validation is conducted over the data and evaluated with the metrics F1, accuracy, positive predictive value, true positive rate and area under the receiver operating characteristic. Results show a generally good performance with a mean F1 of up to 0.900 ± 0.096 using a NN approach. Temporal information indicates low impact on classification performance, while a low-pass filter preprocessing step leads to superior results. Conclusion This work demonstrates the first steps to successfully differentiate healthy brain and tumor tissue using an ultrasonic aspirator during tissue fragmentation. Evaluation shows that both neural network-based classifiers outperform the RF. In addition, the effects of temporal dependencies are found to be reduced when adequate data preprocessing is performed. To ensure subsequent implementation in the clinic, handheld ultrasonic aspirator use needs to be investigated in the future as well as the addition of data to reflect tissue diversity during neurosurgical operations.

Currently, most bone implants used in orthopedics and traumatology are non-degradable and may need to be surgically removed later on e.g. in the case of children. This removal is associated with health risks which could be minimized by using biodegradable implants. Therefore, research on magnesium-based implants is ongoing, which can be objectively quantified through synchrotron radiation microtomography and subsequent image analysis. In order to evaluate the suitability of these materials, e.g. their stability over time, accurate pixelwise segmentations of these high-resolution scans are necessary. The fully-convolutional U-Net architecture achieves a Dice coefficient of 0.750 ± 0.102 when trained with a small dataset with dense expert annotations. However, extending the learning to larger databases would require prohibitive annotation efforts. Hence, in this work we implemented and compared new training methods that require only a small fraction of manually annotated pixels. While directly training on these scribble annotation deteriorates the segmentation quality by 26.8 percentage points, our new random walk-based semi-automatic target achieves the same Dice overlap as a dense supervision, and thus offers a more promising approach for sparse annotations.

This work compares three different approaches to automatically segment the femoral artery from 2D ultrasound images. Two of the architectures follow a sequential structure, where each ultrasound image is considered a slice of the whole vessel volume, and its previous segmentation result will be part of the input, thus leading to a spatial prior. The Dice score on test data show a better performance on the baseline U-Net (0.819) compared to the sequential U-Net approaches (0.633, 0.725) for the femoral artery segmentation. This could be attributed to the misalignment of the slices being used in those networks. A possible improvement could be assumed in the implementation of a spatially calibrated and tracked ultrasound probe. Overall, these results indicate promising approaches for an automatic segmentation of the femoral artery using 2D ultrasound data.