Utilizing Transfer Learning in In Vitro Dissolution Prediction Based on Raman Maps

In vitro dissolution prediction is a field of phar-maceutical research that aims to predict a tablet's dissolution profile using non-destructive methods based on the tablet's measured manufacturing parameters or composition. One of the most popular non-destructive measurements is Raman chemical imaging, which outputs a Raman map. However, because of the current development of the Raman imaging technique, obtaining large training datasets is very expensive and time-consuming. Still, training popular complex deep learning architectures that may achieve better accuracy is only possible with a large dataset. Therefore this article focuses on utilising pre-trained networks and the transfer learning technique to attempt to solve the problem of the small size of training samples. The results showed that deep neural networks with data augmentation and transfer learning could achieve better accuracy than simple convolutional neural networks. The presented strategies showed the need for an extensive dataset could be handled, and the accuracy could be enhanced, although there is room for further refinement.