Application of Deep Learning in 3d Shape Completion (II)

Humans can comprehend and describe the geometry of a complete 3D object based on corrupted data. This project attempts to mimic this ability to reconstruct complete 3D models from incomplete data. On the other hand, due to trauma, congenital defects, plastic surgery, and tumor removal surgery, there is a demand for the repair of skull defects. With the help of an efficient 3D image reconstruction technology, the developed system can help surgeons plan the 3D modeling needed for the medical treatment and benefit patients. In the previous year's project, using the data of stereo images with a resolution of only 64 × 64 × 64 voxels, we developed a primary generative adversarial network that demonstrates promising results. The first year's goal is to develop a deep learning system with 3D shape reconstruction capabilities. The system will be composed of autoencoder and long short-term memory (LSTM) units to meet the requirement of representing high-resolution 3D spatial configuration with contextual details, and use voxel 3D images as the input and output data. In the second year, with the help of two surgeons as co-project inspector, we will verify the feasibility of the 3D medical image restoration system using both simulation and 3D printing technologies. These efforts will help improve its practicability for medical treatments.

Project ID:PB10907-3350
External Project ID:MOST109-2221-E182-025