CR-DM: A novel craniofacial reconstruction framework based on diffusion model

The traditional craniofacial reconstruction methods have limited ability to capture nonlinear changes, which causes challenges in accurately describing the intricate structures between the skull and the craniofacial region. To address this limitation, this paper proposes a novel Craniofacial Reconstruction framework based on Diffusion Model, referred to as CR-DM, in order to perform accurate reconstruction. This model adopts a two-stage approach for craniofacial reconstruction. In the first stage, the encoder and decoder of the pre-trained model are used to compress and reconstruct the feature information of the skull and craniofacial images. In the second stage, bidirectional diffusion between the craniofacial and skull domains is conducted within the latent space. The intermediate feature maps are dynamically adjusted through the feature modulation layer and time-aware encoder, while the lightweight attention mechanism is introduced to enrich the diffusion process. By increasing the scaling factor of specific backbone and skip feature layers, a more flexible and effective denoising effect can be achieved. The proposed method is trained on real craniofacial and skull datasets. The objective evaluation, conducted using several metrics, shows that the proposed model outperforms traditional methods as well as other depth methods. Moreover, ablation experiments are conducted to verify the effectiveness and robustness of the proposed method. The obtained results demonstrate that the proposed method has high craniofacial reconstruction efficiency.