Automated deep learning based detection of cellular deposits on clinically used ECMO membrane lungs

Introduction: Despite the promising application of extracorporeal membrane oxygenation (ECMO) in the treatment of critically ill patients, coagulation-associated technical complications, primarily clot formation and critical bleeding, remain a major challenge during ECMO therapy. The deposition of nucleated cells on the surface has been shown, yet the role of these cells towards complication development is still matter of ongoing research. In particular, the membrane lung (MemL) is prone to clot formation. Therefore, the investigation of nuclear deposits on its hollow-fibers may provide insights for a better understanding of the cellular mechanisms involved in the development of ECMO complications. Methods: To support current research, this study aimed to develop a deep learning-based tool for the automated detection and quantitative analysis of nuclear depositions on MemL hollow-fiber mats. A customized fluorescence microscopy workflow, combined with a semi-automated iterative labeling strategy, was used to generate a high-quality dataset for model training. Results: Six configurations of instance segmentation models were evaluated, with a Mask R-CNN with ResNet 101 backbone using dilated convolution providing the most balanced performance in both nuclei count and area accuracy. Compared with U-Net-based approaches such as Cellpose or StarDist, the proposed model demonstrated superior segmentation of overlapping and low-intensity nuclei, maintaining accuracy even in densely packed cellular regions. Discussion: We present an automated image analysis tool for clinically used MemLs, which exhibit complex three-dimensional hollow-fiber architectures and irregular cellular deposits that challenge conventional tools. A dedicated graphical user interface enables streamlined detection, morphometric analysis, and spatial clustering of nuclei, establishing a reproducible workflow for high-throughput analysis of fluorescence microscopy images. This approach eliminates labor-intensive manual counting and facilitates large-scale studies on cell-fiber interactions and disease-related correlations.