Accurate 3D reconstruction of the defective part is critically important for repairing defects in the human skull. After investigating the feasibility of 3D convolution neural network (CNN)-based approach, DeepMedic CNN is chosen for repairing defects of the human skull. Training set of 3D CNN model is produced by randomly segmenting the initial 3D model of the skull which come from a whole CT scan of a healthy person. The 3D CNN model was evaluated using a computer-simulated 3D skull model containing the defective part, and in vivo patient. The results showed that based on 160 groups of computer-simulated 3D CT data, the average dice similarity coefficient (DSC), sensitivity (SE) and Hausdorff distance (HD) are 89.31%, 91.81%, and 25.9%, respectively. These quantitative indexes showed that the proposed method is able to do a reliable bone structure predication. For in vivo patient, the obtained model is also able to generate a suitable 3D bone model for the data under consideration. This approach could increase the computational efficiency of the repairing process without the need for segmentation and reconstruction of the skull, and thereby has potential applications to motivating further accurate repairing of defects of skull.

中文翻译：

---

基于神经网络的颅骨缺陷修复：性能和可行性的初步评估

缺陷部分的准确 3D 重建对于修复人类头骨中的缺陷至关重要。在研究了基于 3D 卷积神经网络 (CNN) 的方法的可行性后，DeepMedic CNN 被选择用于修复人类头骨的缺陷。3D CNN 模型的训练集是通过随机分割来自健康人的整个 CT 扫描的头骨的初始 3D 模型而产生的。使用包含缺陷部分的计算机模拟 3D 颅骨模型评估 3D CNN 模型，并且体内病人。结果表明，基于160组计算机模拟3D CT数据，平均骰子相似系数（DSC）、灵敏度（SE）和豪斯多夫距离（HD）分别为89.31%、91.81%和25.9%。这些定量指标表明，所提出的方法能够进行可靠的骨结构预测。为了体内患者，获得的模型还能够为所考虑的数据生成合适的 3D 骨骼模型。这种方法可以提高修复过程的计算效率，而不需要对颅骨进行分割和重建，从而在促进颅骨缺损的进一步准确修复方面具有潜在的应用价值。