Automatic Cranial Defect Reconstruction with Self-Supervised Deep Deformable Masked Autoencoders.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Cranial injuries necessitate personalized implants, requiring time-consuming manual design and manufacturing.
• Automating cranial defect reconstruction is crucial for efficiency and cost reduction.
• Current deep learning methods for volumetric shape completion often rely on costly, time-consuming manual or synthetic ground-truth annotation.

Purpose of the Study:

• To propose a novel, efficient, and automated approach for cranial defect reconstruction.
• To overcome the limitations of supervised learning methods in defect annotation for cranial implants.
• To enhance the generalizability of deep learning models for volumetric shape completion tasks.

Main Methods:

• Utilized a self-supervised masked autoencoder for volumetric shape completion.
• Trained the model on the SkullBreak and SkullFix datasets.
• Compared the proposed method against several state-of-the-art deep neural networks.

Main Results:

• The self-supervised masked autoencoder demonstrated quantitative and qualitative improvements over existing methods.
• The approach inherently increases training set heterogeneity, acting as a form of data augmentation.
• Achieved efficient real-time reconstruction of cranial defects.

Conclusions:

• Self-supervised masked autoencoders offer a viable and effective alternative for automated cranial defect reconstruction.
• The proposed method enhances data heterogeneity, leading to improved model generalizability.
• This technique enables efficient, real-time cranial defect repair, advancing personalized implant solutions.