Quantification of tumorsphere migration with a physics-based machine learning method.

Current analysis techniques available for migration assays only provide quantitative measurements for overall migration. However, the potential of regional migration analyses can open further insight into migration patterns and more avenues of experimentation with the same assays. Previously, we developed an analysis pipeline utilizing the finite element (FE) method to show its potential in analyzing glioblastoma (GBM) tumorsphere migration, especially in characterizing regional changes in the migration pattern. This study aims to streamline and further automate the analysis system by integrating the machine-learning-based U-Net segmentation with the FE method. Our U-Net-based segmentation achieved a 98% accuracy in segmenting our tumorspheres. From the segmentations, FE models made up of 3D hexahedral elements were generated, and the migration patterns of the tumorspheres were analyzed under treatments B and C (under non-disclosure agreements). Our results show that our overall migration analysis correlated very strongly (R of 0.9611 and 0.9986 for treatments B and C, respectively) with ImageJ's method of migration area analysis, which is the most common method of tumorsphere migration analysis. Additionally, we were able to quantitatively represent the regional migration patterns in our FE models, which the methods purely based on segmentations could not do. Moreover, the new pipeline improved the efficiency and accessibility of the initial pipeline by implementing machine learning-based automated segmentation onto a mainly open-sourced FE analysis platform. In conclusion, our algorithm enables the development of a high-content and high-throughput in vitro screening platform to elucidate anti-migratory molecules that may reduce the invasiveness of these malignant tumors.