Computer vision and hybrid machine learning algorithms on glioblastoma 3D cell culture multiplex environment

Artificial intelligence (AI) is primarily comprised of machine learning and deep learning computational methodologies, offering substantial promise for improving the therapeutic management of glioblastoma, the most prevalent and aggressive adult brain tumor in the central nervous system. The objective of this study was to predict the migration and proliferation behavior of U87 human glioblastoma cells interacting with I-NHA human astrocytes. These engineered cell models, embedded within specific hydrogels, enabled us to mimic the organisms and observe their interactions. Our research involved a significant amount of raw data obtained from classical software calculations combined with image datasets captured from three-dimensional (3D) tumor microenvironment studies. In vitro studies indicate that glioblastoma exhibits distinct cellular characteristics, including migration, invasion, and proliferation behaviors. In this context, this work evaluated the effectiveness of Recurrent Neural Network (RNN) algorithms on 300 normalized image data obtained from two technical replicates and calculated the corresponding parameters as AI features. Specifically, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models of RNN, along with their hybrid variants with Convolutional Neural Network (CNN), were utilized to apply machine learning techniques to cell culture images. Among these models, GRU demonstrated superior estimation performance, particularly when the data density decreased; the optimized GRU model achieved a maximum accuracy of 0.85. While transformer-based models are highly effective for processing large-scale datasets due to their capacity to encode long-range dependencies, RNN-based approaches were observed to outperform transformers in time-series analyses involving medium-sized cell image datasets. These findings emphasize that enhancing image data quality and augmenting datasets with both technical and biological replicates could enhance the performance of machine learning algorithms. In summary, while this investigation demonstrates encouraging predictive accuracy, it serves primarily as a proof of concept. AI-based analyses in glioblastoma research have demonstrated promising results; further investigation and integration of larger and more diverse data are required to standardize methodology and improve the interpretability and generalizability of the prediction results.