Automation of Decellularization Process Using Artificial Neural Networks

A neural network architecture inspired by AlexNet was created and adapted for regression, which predicts continuous values instead of classifying images. The network was trained on images with a resolution of
$\mathbf{1024}\times \mathbf{768}$
pixels, taken during the decellularization processes. This database was created along with the experiments carried out within the OncoGen Research Institute, from Tlmişoara, using the modified Langerdorff device for rat heart decellularization. The biochemical values, DNA, and protein concentrations were measured at 30-minute intervals using a micro-spectrophotometer and correlated with the corresponding times of the acquired images. The model was trained on 46.938 images using data augmentation techniques such as rotation, flipping, and zooming. The model was tested on a separate set of 2000 images, and the performance was evaluated by comparing the predicted values of DNA and protein concentrations to the actual values. The created convolutional neural network architecture is efficient in detecting and predicting continuous values for DNA and protein concentration.