Cleft Volume Estimation and Maxilla Completion Using Cascaded Deep Neural Networks

In this paper, we propose an end-to-end cascaded deep neural network based-framework for the prediction of cleft volume and maxilla completion in the alveolar cleft grafting procedures. We devise the coupled cascaded deformable volumetric registration and cleft prediction networks with progressively refined cleft masks. The framework can be stacked on an existing volumetric registration network for partial registration between the template volume with the complete maxilla and the one with cleft lips and palates (CLP). Instead of one-shot registration-based volume completion for the cleft volume prediction, we present a cascaded registration network to accommodate coarse-to-fine volumetric transformations, enabling the refinement of the cleft volume and fine-tuning of cleft prediction network. The resulting dense displacement fields facilitate the cleft defect location and virtual maxilla completion. The iteratively updated cleft volume from the partial registration is utilized to refine the end-to-end cleft prediction network, which avoids the Boolean operation-based cleft estimation in the online testing process. We devise an alternating optimization approach to fine-tune the registration and cleft prediction networks. Qualitative and quantitative comparisons of the proposed approach on clinically-obtained CLP CBCT images demonstrate that our method is effective for cleft volume estimation and virtual maxilla completion.