(C-95) A suite of novel automated image analysis tools for high-throughput visualization and assessment of 3D tumor models

Introduction:: 3D tumor models have emerged as essential tools for studying tumor progression, invasion, and angiogenesis in vitro. However, there are currently limited open source options for automated assessment of these 3D cell behaviors. To address this challenge, we developed a Python-based open-source software package for the high-throughput analysis of cancer and endothelial cell dynamics in hydrogels. We applied this software to 3D multilayer multicellular models of cervical and endometrial cancer cells co-cultured with human microvascular endothelial cells (hMVEC). Our software package automates the quantification of cell coverage area, invasion depth, and microvessel formation, enabling rapid and accurate assessment of phenotypic cell responses in 3D tumor models.

Materials and Methods:: Multilayer hydrogels were fabricated following established protocols, with hMVEC cells and cancer cells (cervical: SiHa, CaSki; endometrial: HEC-1A) cocultured in 96-well plates. The phenotypic response was defined as cancer cell coverage area, cancer cell invasion, endothelial cell coverage area, and endothelial microvessel length. A Cytation 5 V3.14 cell imaging multimode reader was used to acquire two-channel 85 µm z-stack images every three hours for 48 hours.

Cell area was calculated by using focus stacking to compress the Z-stack into 2D images followed by a Gaussian mixture model-based technique to classify background and foreground pixels. These images can then interface with FIJI (NIH) to measure cell area, shape, and orientation. To calculate cell invasion depth and microvessel formation we built two deep learning systems that iterate test and training datasets to improve accuracy over time. Specifically, invasion depth was calculated by applying a ResNet50-based classification model to identify the deepest point of invasion for any given z-stack. Microvessel formation was calculated by combining a U-Net Xception-style segmentation model with a graph reconstruction algorithm to extract a representation of the microvessel network. We then applied a persistent homology filtration to quantify the branching structure of microvessel formation (Fig. 1).

Results, Conclusions, and Discussions:: Results
To evaluate the accuracy of our software, we compared the automated measurements of cell coverage, invasion depth, and total microvessel length with manual measurements. The coefficient of determination (R2) was approximately 0.9 for cell coverage, 0.98 for invasion depth, and 0.8 for microvessel length (Fig. 2). These findings demonstrate that the developed software package is a reliable tool for high-throughput quantification of phenotypic cell responses in 3D tumor models.

Conclusion
We have developed an open-source software package that enables researchers to automate image analysis of 3D tumor models. This software can be integrated into existing workflows and adapted for particular datasets of 3D in vitro models, facilitating rapid and accurate assessment of tumor progression, invasion, and angiogenesis in diverse cancer types.

Acknowledgements (Optional): :

References (Optional): :