Comparative evaluation of neural networks and transfer learning for predicting mechanical properties of 3D‐printed bone scaffolds

Due to the complexity of bone structures, influenced by factors such as injuries, defects, age, and health conditions, creating accurate 3D personalized scaffolds is challenging and requires extensive trial‐and‐error. To ensure high‐quality bespoke scaffolds, three key mechanical parameters—volume fraction, Poisson's ratio, and elastic modulus—must be accurately calculated during surgery. This research proposed the idea of applying convolutional neural networks (CNNs) and Transfer Learning (TL) to predict the volume fraction, Poisson's ratio and elastic modulus of medical bone scaffolds. Deep neural networks require extensive training data, but obtaining a large volume of labelled data is often challenging. In this study, a comprehensive dataset is generated by the parametric implicit equation of Body Centered Cubic (BCC) lattice structure to train deep neural networks. The result shows that the CNNs adopt better in comparison to Transfer Learning Resnet‐50, Resnet‐15 and Resnet‐34 for predicting the fundamental properties of medical bone implants. The Resnet‐50 model demonstrates the highest MEA (Mean Absolute Error) in volume fraction (1.382), Poisson's ratio (0.006), and elastic modulus (12.551), while the CNN model shows the lowest values, with volume fraction MEA of 0.348, Poisson's ratio of 0.003, and elastic modulus of 4.853. The research findings can be used by medical surgeons and biomechanical engineers to calculate the mechanical properties of 3D scaffolds.