ScaffoldGAN: Synthesis of Scaffold Materials based on Generative Adversarial Networks

Abstract Digitally synthesizing scaffold-like materials with complex structures, e.g., bones or metal foam, is a fundamental yet challenging task in tissue engineering and other biomedical applications, because it is difficult to generate synthesized results with equal visual complexity, strong spatial coherence, and similar statistical metrics. To handle these challenges, we present ScaffoldGAN, an efficient end-to-end framework based on generative adversarial networks (GANs) for synthesizing three-dimensional (3D) materials with complex internal structures resembling the given exemplar. Specifically, we propose a novel structural loss to enforce strong spatial coherence in the synthesized results by leveraging the deep features learned by our networks. To demonstrate the effectiveness of our model and the proposed structural loss term, we collected example data containing various structural complexities, covering two categories of materials, i.e., bones and metal foams. Extensive comparative experiments on these collected data showed that our method outperforms state-of-the-art methods, producing synthesized results with better visual quality and desirable statistical metrics. The ablation study proves the structural loss is the main contributor to the performance gain, validating our design choice.