(C-86) An Artificial Intelligence Software for Automatically Analyzing the Interfacial Spheroid Invasion in Tumor Microenvironment using Differential Interference Contrast Microscopic Images

Introduction:: Interfacial invasion is a complex process in cancer metastasis that involves the interaction between cancer cells and the extracellular matrix (ECM) in tumor microenvironment. Understanding its mechanisms is critical for developing new therapeutic strategies. The spheroid invasion model has the potential to be the preferred model for investigating interfacial invasion since it can mimic the 3D structure and microenvironment of tumors in vivo, providing a more physiologically relevant platform. To analyze the interfacial spheroid invasion model, differential interference contrast (DIC) microscopy is commonly used. However, analyzing spheroid DIC images is challenging and time-consuming, making it difficult to extract and analyze the model objectively.


Materials and Methods:: A software was developed using a deep learning-based method with a Convolutional Neural Network (CNN) to segment spheroid core and invasive area from DIC images, employing EfficientNet as the encoder and Unet++ as the decoder. For single analysis, the CNN combines RegNetY320 as the encoder and Feature Pyramid Network (FPN) as the decoder. The model was trained on 2795 images, split into training and validation sets, with pre-trained weights, regularization, and evaluated using the Intersection over Union (IoU) metric. The software calculates quantitative parameters through OpenCV functions after post-processing.


Results, Conclusions, and Discussions:: We present 5-fold cross-validation results for Intersection over Union (IoU) at different thresholds for the invasive area, spheroid core, and single cell models. The EfficientNet-B7 and Unet++ (256, 128, 128, 64, 64) architecture achieved the highest IoU for the invasive area and spheroid core models, with an average IoU of 0.77 and 0.75, respectively. For single cell segmentation, the RegNetY320 encoder and FPN decoder model achieved an IoU of 0.48, an improvement of 17.5% compared to EfficientNet-B7 and Unet++. Our automated segmentation software successfully characterizes the influence of stiffness and interfacial adhesive strength on interfacial spheroid invasiveness, defining area, radius, perimeter, and complexities parameters. By applying the software, significantly different of spheroid invasion patterns in different interface conditions can be found in their early stage of invasion from 8 to 10 hours.

A deep learning-based software was developed to automatically segments spheroid core, invasive area, and single cell from DIC images, enabling the measurement and analysis of spheroid migration dynamics on varying interface conditions. Our software offers a fluorescence-free and non-destructive approach, which can be useful for assessing metastatic activity in drug screening research with reduced waiting times. The deep learning model of our software has the potential to expand its segmentation capability for various cancer cell types in the future.


Acknowledgements (Optional): :

References (Optional): :