Conference Agenda

The lack of labeled, intraoperative patient data in medical scenarios poses a relevant challenge for machine learning applications. Given the apparent power of machine learning, this study examines how synthetically-generated data can help to reduce the amount of clinical data needed for robust liver surface segmentation in laparoscopic images. Here, we report the results of three experiments, using 525 annotated clinical images from 5 patients alongside 20,000 synthetic photo-realistic images from 10 patient models. The effectiveness of the use of synthetic data is compared to the use of data augmentation, a traditional performance-enhancing technique. For training, a supervised approach employing the U-Net architecture was chosen. The results of these experiments show a progressive increase in accuracy. Our base experiment on clinical data yielded an F1 score of 0.72. Applying data augmentation to this model increased the F1 score to 0.76. Our model pre-trained on synthetic data and fine-tuned with augmented data achieved an F1 score of 0.80, a 4% increase. Additionally, a model evaluation involving k-fold cross validation highlighted the dependency of the result on the test set. These results demonstrate that leveraging synthetic data has the ability of limiting the need for more patient data to increase the segmentation performance.

Core-shell capsules (CSC) are a promising approach for 3D cell culture because they overcome the challenges of traditional large-scale cell cultivation techniques used in tissue engineering. Currently, CSC are segmented from microscopic images in a cumbersome manual procedure to evaluate their properties, such as size or complete encapsulation of the core compartment.

In this paper, we propose an automated segmentation process of CSC based on an unmodified YOLOv8 instance segmentation model. We train the model exclusively on synthetic CSC images created from 10 manually annotated real images and evaluate its performance using the common Intersection over Union (IoU) metric on a test set consisting of 181 real images. Without modifying the model or tuning the hyperparameters, we achieve a mean IoU of 0.86, underlining the potential of deep-learning-based CSC segmentation relying entirely on synthetic training data.

Human activity recognition (HAR) is an expanding area of research. Although sensors are becoming more readily available, there is a trend toward minimizing the number of associated sensors for better applicability. Neural networks are often employed for HAR, as they are capable of identifying movement patterns within data. To optimize the amount of useful information provided to the network, feature extraction methods are commonly applied. However, these feature extraction methods are time-consuming and, thus, not applicable to real-time applications. In this work, we investigate the impact of using quaternions as input features for an LSTM network on HAR, specifically focusing on level walking and stair climbing activities, while also considering the inference time. We demonstrate that raw IMU data, i.e., acceleration and gyroscope combined with quaternions significantly enhance classification accuracy without adversely affecting the inference time.

Minimally invasive transcatheter aortic valve replace-ment (TAVR) procedure has become the preferred proce-dure for patients with aortic valve stenosis or insufficiency with high risk for conventional open surgery. The favora-ble clinical outcomes of high-risk patients led to an ex-pansion of the cohort including intermediate and low-risk patients. A critical aspect of advancing TAVR procedures lies in preoperative planning, integrating patient-specific in-silico deployment simulation and post-deployment fluid mechanics assessments. This study introduces a novel approach to calcified TAVR patient shape model-ing, addressing this problems. The model integrates an extended mesh generated by DeepCarve, encompassing the aortic arch, and a novel deep learning-based volumet-ric shape model of calcifications. The key innovation lies in the utilization of a conditional Convolutional Varia-tional Autoencoder (cCVAE) to generate realistic calcifi-cation patterns, demonstrating promising preliminary results in matching actual cohort data. Future investiga-tions should focus on data collection from diverse medi-cal centers to validate and refine the proposed methodol-ogy. This study showcases significant progress in generat-ing synthetic TAVR patient geometries, incorporating detailed anatomical structures such as the aortic root, valve, and arch, along with volumetric calcification pat-terns. These findings represent a crucial step towards ena-bling real-time preoperative TAVR planning, inclusive of patient-specific in-silico deployment simulation and comprehensive fluid mechanics assessments.

This study presents a novel approach to augment physiotherapeutic exercise datasets by synthesizing realistic inertial measurement unit (IMU) data. The augmented dataset is used to improve the performance of a deep learning based exercise evaluation system. The approach is demonstrated using the deep squat exercise from the Functional Movement Screening (FMS) protocol. By integrating musculoskeletal simulation and leveraging knowledge of potential movement errors based on FMS evaluation criteria, we aim to produce synthetic data that closely mimics human movement. Our evaluation demonstrates that training a CNN-LSTM neural network with both real and synthesized data significantly improves the model's performance, especially in generalizing to unknown subjects. However, limitations such as the approach's specificity to the deep squat exercise suggest the need for a more adaptable method. Future work will focus on refining the synthesis process to ensure a broader applicability to various exercises. This research contributes to advancing automated physiotherapeutic exercise evaluation, highlighting the importance of synthetic data in achieving better performing and more generalizable models.