InceptMan: An InceptionNeXt-Based Architecture for End-to-End Mandible Reconstruction

Mandible reconstruction is one of the virtual surgical planning (VSP) steps in mandible reconstruction surgery, particularly in cases involving tumors or radical cancer surgery. While deep learning approaches have shown success in medical image reconstruction, existing methods struggle with the complex morphology of mandible structures and often require manual intervention for defining resection margins. To address these challenges, we present a novel UNet-based architecture based on CraNeXt and InceptionNeXt designed for automated mandible reconstruction, InceptMan. In addition, we proposed the automated cutting pipeline to define resection margins between healthy and defective parts of the mandible before reconstruction. We evaluated our proposed model and pipeline on two datasets: patient data with mandible abnormalities in the unhealthy dataset and healthy controls in the healthy dataset. On the healthy dataset, InceptMan outperformed the baseline 3DRDUNet, improving the dice similarity coefficient (DSC) from 0.7125 to 0.8151 with performance comparable to the state-of-the-art CraNeXt, while also achieving $1.4\times $ higher training throughput. On the unhealthy dataset, evaluation with our proposed metrics, Surface Recall and $95^{\text {th}}$ percentile Asymmetric Hausdorff distance (dAH95), showed that the automated cutting pipeline performs comparably to manual cutting. These results highlight the pipeline’s potential for efficient and accurate automated mandible reconstruction in clinical applications.