A deep learning approach for automated inspection of 3D printed orodispersible films

Abstract. This work aims at demonstrating the applicability of additive manufacturing and machine vision technologies for the on-demand manufacture of customized medicinal products at the point of need, while conforming to fundamental quality guidelines such as cGood Manufacturing Practices (cGMPs), Quality by Design (QbD) as well as Quality Risk Management (QRM) and assuring product quality. The adopted approach combines a 3D printing process for orodispersible films and a deep neural network model in order to develop an automated quality control system for defect detection. The orodispersible films were manufactured using an off-the-shelf 3D printer equipped with a modified semisolids printhead, with Warfarin Sodium (C19H15NaO4) selected as the active pharmaceutical ingredient. The dose could be accurately regulated by varying the rectangular film’s dimensions. Three different defect types were considered, corresponding to layer shifting, incomplete prints and inclusion of bubbles within the film, respectively. A pretrained convolutional neural network (GoogleNet) was fine-tuned to perform the defect classification of the 3D printed films, which after data augmentation and hyperparameter tuning, achieved an accuracy of up to 100% in detecting conforming and not conforming products.