2026-07-21 –, Room 1.19 (Ground Floor, Shannon)
The transition toward Organ-on-Chip (OoC) and Engineered Heart Tissues (EHTs) in pharmacological research necessitates automated tools for longitudinal cell monitoring. Manual tracking of cardiac cycles is labor-intensive and impractical for long-duration studies, highlighting the need for high-throughput segmentation. This study proposes a specialized shallow 2D U-Net architecture utilizing Focal Loss to effectively manage class imbalances between cellular structures and the background in spatial recordings.
The model was trained on a dataset of 1,240 images and validated against 100 test samples, achieving a Dice score of 0.9152 and a mean Intersection over Union (mIoU) of 0.8486 relative to manual ground truths. To contextualize these results, the proposed network was benchmarked against MedSAM, a state-of-the-art foundation model for medical imaging.
While MedSAM offers broad generalization, our lightweight U-Net demonstrates superior computational efficiency and localized precision for the high-frame-rate requirements of cardiac cycle analysis. The findings suggest that a task-specific, loss-optimized architecture provides a high-performance, resource-efficient alternative to large-scale foundation models, facilitating objective and scalable analysis of single-cell behavior in animal-free drug testing environments.
The paradigm shift in pharmacological research towards Organ-on-Chip (OoC) and Engineered Heart Tissues (EHTs) has significantly reduced reliance on animal models by providing physiologically relevant simulations of human cardiac environments. However, the longitudinal observation of single-cell dynamics during the cardiac cycle, particularly following the administration of novel therapeutic agent, presents a substantial computational challenge. Manual tracking of morphological transitions over extended durations is labor-intensive and prone to observer bias, necessitating high-throughput, fully automated segmentation frameworks.
In this study, we propose an optimized, shallow 2D U-Net architecture specifically tailored for the segmentation of EHT spatial recordings. To address the inherent class imbalance between the cellular foreground and the interstitial background, we implemented Focal Loss during the training phase, forcing the network to prioritize hard-to-classify edge pixels. The model was trained on a curated dataset of 1,240 high-resolution images and validated against a blind test set of 100 samples.
Our proposed lightweight model achieved a Dice similarity coefficient of 0.9152 and a mean Intersection over Union (mIoU) of 0.8486 relative to ground-truth manual segmentations. To contextualize these results, we performed a comparative benchmark against MedSAM, a state-of-the-art foundation model for medical image segmentation. While MedSAM demonstrates high generalization, our results indicate that the task-specific shallow U-Net offers superior computational efficiency and localized accuracy for the rapid frame rates required in EHT cardiac cycle analysis. This work demonstrates that an optimized, specialized architecture can match or exceed foundation model performance in niche biomedical applications while remaining viable for real-time monitoring.
Maciej Szymkowski is a Ph.D. in the field of Computer Science. From April 2025, he is working with Future Processing as a Senior Machine Learning Engineer. Previously he was with Bialystok University of Technology (Faculty of Computer Science, 2018-2026), Warsaw University of Technology (Faculty of Electronics and Information Technology, 2021-2022) and AGH in Cracow (Faculty of Physics and Applied Computer Science, 2021-2022). He gained his experiences also as a researcher in diversified projects and software developer/engineer in different companies (e.g., SoftServe, Symmetra, Transition Technologies, Hemolens Diagnostics). He was also with Łukasiewicz Research Network - Poznan Institute of Technology (2022-2024) where he was a Head of AI Development section. Maciej Szymkowski is an author or co-author of more than 45 research papers published in JCR journals, as chapters in the books or in international conference proceedings. His main research area is Computer Vision – especially in the field of medicine and transport. He is interested in Vision-Language Models (VLMs), Large Language Models (LLMs) and machine learning algorithms. In his free time, he loves to extend his knowledge, take a long walk, read a book or watch soccer, basketball or hockey. He is a fan of Legia Warsaw, Real Madrid (soccer), New York Knicks (basketball) and Pittsburgh Penguins (hockey).