Neural Network Development for Explainability Analysis of Cell Behaviors in Co-culture Experiments

This thesis investigates the classification of cells in in-vitro co-culture settings using explainable deep learning models. Confocal microscopy videos, provided by the Dipartimento Rizzoli-RIT (Istituto Ortopedico Rizzoli), are processed to build a robust dataset including 143b osteosarcoma and Mesenchymal Stem Cells (MSCs). Segmentation errors and missing data are addressed through automated Python tools combining contour extraction, bounding box reconstruction, and linear interpolation, and a rotation-based strategy is implemented for data augmentation. Two distinct classification paradigms are investigated. The first one adopts a static framework based on a Multilayer Perceptron (MLP) architecture, where each cell at a specific time frame is an independent sample. Architectural and training hyperparameters are optimized through the Optuna framework. The second approach aims to capture temporal dependencies in cellular trajectories, a sequential approach is implemented by using a Long Short-Term Memory (LSTM ) architecture. Model interpretability is ensured by using SHAP for MLP and TimeSHAP for LSTM, thus attributing prediction relevance to shape-based features and temporally pruned events, respectively. The models achieve over 94% accuracy in the static setting and approximately 87% in the temporal case. Results confirm that biologically meaningful descriptors — such as Circularity, Area, and Convexity — dominate the classification decision