Core Concepts
A self-supervised deep deformable masked autoencoder approach can efficiently reconstruct cranial defects, outperforming state-of-the-art supervised segmentation methods.
Abstract
The paper proposes an alternative approach to automatic cranial defect reconstruction using a deep deformable masked autoencoder. The key highlights are:
The method uses a self-supervised approach, where random patches are deformably masked from healthy skull inputs to generate defective samples. A deep encoder-decoder network is then trained to reconstruct the missing parts.
This self-supervised approach increases the heterogeneity of the training data, leading to improved generalization compared to supervised segmentation methods that rely on synthetic defect generation.
The proposed method outperforms several state-of-the-art supervised segmentation networks, such as Residual UNet, UNETR, and SwinUNETR, on the SkullBreak and SkullFix datasets. It achieves higher Dice scores, better boundary Dice, and lower Hausdorff distances.
The authors show that the random elastic deformations applied to the masked patches are crucial for improving performance, especially on the SkullBreak dataset with smooth, irregular defects.
The method can be easily extended to new datasets containing healthy skulls, without the need for manual preprocessing and defect synthesis.
The main drawback is the longer training time compared to supervised segmentation approaches, but the inference time remains comparable.
Overall, the proposed self-supervised deep deformable masked autoencoder provides an efficient and generalizable solution for automatic cranial defect reconstruction.
Stats
Thousands of people suffer from cranial injuries every year.
The SkullFix and SkullBreak datasets used for training and evaluation contain a total of 214 and 570 cases, respectively.
The proposed method improves the Dice score by more than 0.1 and the Hausdorff distance by 1.0 mm compared to state-of-the-art segmentation networks.
Quotes
"The task is considerably more difficult than training the classical autoencoders to just recover its input. The self-supervisedly pretrained masked autoencoders are then useful for other downstream tasks because they learn both general and detailed features associated with the data."
"Since the processed data is binary, we reconstruct the patches using the Soft Dice Score as the loss function. The initial ablations confirmed that such an approach is more stable and converges faster than experiments using mean absolute or mean squared differences."