Deep Learning for Cranial Defect Reconstruction

Jianning Li

Research output: ThesisMaster's Thesis


A fast and fully automatic design of 3D printed patient-specific cranial implants is highly desired in cranioplasty, a process to restore a defect on the skull. We formulate skull defect restoration as a 3D volumetric shape completion task, where a partial skull volume is completed automatically. The difference between the completed skull and the partial skull is the restored defect, in other words, the implant that can be used in cranioplasty. To this end, a deep neural network based on the encoder-decoder architecture is adopted. To facilitate supervised training, a database containing 167 healthy skulls segmented from head CT in clinical routine has been established. Synthetic defects are injected into the healthy skull to create training and evaluation data pairs. To work on high-resolution and spatially sparse skull data, we proposed a tailored patch-based training scheme that overcomes the disadvantages of conventional patch-based training method and shows significant improvement. In particular, the patch-based training method is applied to images of high resolution and proves to be effective in tasks such as segmentation. However, we demonstrate that conventional patch-based training method is suboptimal for tasks such as shape reconstruction, where the overall shape distribution of the target has to be learnt, since if cannot be captured efficiently by a sub-volume cropped from the target. Additionally, the standard dense implementation of a convolutional neural network (CNN) tends to performs poorly on sparse data such as the skull which has a low voxel occupancy rate. Our tailored training scheme encourages a standard CNN to learn interpretable features from the high-resolution and sparse data. We have evaluated our method on both skulls with synthetic defects and skulls with real defects manually injected by neurosurgeons in craniotomy, and the results show potential for clinical applicability.
Original languageEnglish
QualificationMaster of Science
Awarding Institution
  • Graz University of Technology (90000)
  • Egger, Jan, Supervisor
  • Schmalstieg, Dieter, Supervisor
Publication statusPublished - Jan 2020


  • cranial implant
  • cranioplasty
  • craniotomy
  • deep learning
  • shape completion
  • skull database
  • skull reconstruction

ASJC Scopus subject areas

  • Artificial Intelligence
  • Computer Vision and Pattern Recognition
  • Computer Graphics and Computer-Aided Design

Fields of Expertise

  • Information, Communication & Computing

Cite this