Virtual reality for assessing stereopsis performance and eye characteristics in Post-COVID.

In 2019, we faced a pandemic due to the coronavirus disease (COVID-19), with millions of confirmed cases and reported deaths. Even in recovered patients, symptoms can be persistent over weeks, termed Post-COVID. In addition to common symptoms of fatigue, muscle weakness, and cognitive impairments, visual impairments have been reported. Automatic classification of COVID and Post-COVID is researched based on blood samples and radiation-based procedures, among others. However, a symptom-oriented assessment for visual impairments is still missing. Thus, we propose a Virtual Reality environment in which stereoscopic stimuli are displayed to test the patient's stereopsis performance. While performing the visual tasks, the eyes' gaze and pupil diameter are recorded. We collected data from 15 controls and 20 Post-COVID patients in a study. Therefrom, we extracted features of three main data groups, stereopsis performance, pupil diameter, and gaze behavior, and trained various classifiers. The Random Forest classifier achieved the best result with 71% accuracy. The recorded data support the classification result showing worse stereopsis performance and eye movement alterations in Post-COVID. There are limitations in the study design, comprising a small sample size and the use of an eye tracking system.

Wearable Sensors for Activity Recognition in Ultimate Frisbee Using Convolutional Neural Networks and Transfer Learning.

In human activity recognition (HAR), activities are automatically recognized and classified from a continuous stream of input sensor data. Although the scientific community has developed multiple approaches for various sports in recent years, marginal sports are rarely considered. These approaches cannot directly be applied to marginal sports, where available data are sparse and costly to acquire. Thus, we recorded and annotated inertial measurement unit (IMU) data containing different types of Ultimate Frisbee throws to investigate whether Convolutional Neural Networks (CNNs) and transfer learning can solve this. The relevant actions were automatically detected and were classified using a CNN. The proposed pipeline reaches an accuracy of 66.6%, distinguishing between nine different fine-grained classes. For the classification of the three basic throwing techniques, we achieve an accuracy of 89.9%. Furthermore, the results were compared to a transfer learning-based approach using a beach volleyball dataset as the source. Even if transfer learning could not improve the classification accuracy, the training time was significantly reduced. Finally, the effect of transfer learning on a reduced dataset, i.e., without data augmentations, is analyzed. While having the same number of training subjects, using the pre-trained weights improves the generalization capabilities of the network, i.e., increasing the accuracy and F1 score. This shows that transfer learning can be beneficial, especially when dealing with small datasets, as in marginal sports, and therefore, can improve the tracking of marginal sports.

From the Laboratory to the Field: IMU-Based Shot and Pass Detection in Football Training and Game Scenarios Using Deep Learning.

The applicability of sensor-based human activity recognition in sports has been repeatedly shown for laboratory settings. However, the transferability to real-world scenarios cannot be granted due to limitations on data and evaluation methods. On the example of football shot and pass detection against a null class we explore the influence of those factors for real-world event classification in field sports. For this purpose we compare the performance of an established Support Vector Machine (SVM) for laboratory settings from literature to the performance in three evaluation scenarios gradually evolving from laboratory settings to real-world scenarios. In addition, three different types of neural networks, namely a convolutional neural net (CNN), a long short term memory net (LSTM) and a convolutional LSTM (convLSTM) are compared. Results indicate that the SVM is not able to reliably solve the investigated three-class problem. In contrast, all deep learning models reach high classification scores showing the general feasibility of event detection in real-world sports scenarios using deep learning. The maximum performance with a weighted f1-score of 0.93 was reported by the CNN. The study provides valuable insights for sports assessment under practically relevant conditions. In particular, it shows that (1) the discriminative power of established features needs to be reevaluated when real-world conditions are assessed, (2) the selection of an appropriate dataset and evaluation method are both required to evaluate real-world applicability and (3) deep learning-based methods yield promising results for real-world HAR in sports despite high variations in the execution of activities.

Deep learning-based detection of motion artifacts in probe-based confocal laser endomicroscopy images.

PURPOSE: Probe-based confocal laser endomicroscopy (pCLE) is a subcellular in vivo imaging technique capable of producing images that enable diagnosis of malign structural modifications in epithelial tissue. Images acquired with pCLE are, however, often tainted by significant artifacts that impair diagnosis. This is especially detrimental for automated image analysis, which is why said images are often excluded from recognition pipelines. METHODS: We present an approach for the automatic detection of motion artifacts in pCLE images and apply this methodology to a data set of 15 thousand images of epithelial tissue acquired in the oral cavity and the vocal folds. The approach is based on transfer learning from intermediate endpoints within a pre-trained Inception v3 network with tailored preprocessing. For detection within the non-rectangular pCLE images, we perform pooling within the activation maps of the network and evaluate this at different network depths. RESULTS: We achieved area under the ROC curve values of 0.92 with the proposed method, compared to 0.80 for the best feature-based machine learning approach. Our overall accuracy with the presented approach is 94.8%. CONCLUSION: Over traditional machine learning approaches with state-of-the-art features, we achieved significantly improved overall performance.