基於二維X光影像之三維膝蓋骨骼模型重建

3D Knee Bone Reconstruction from 2D X-Ray Images

賴岳群(Yueh-Chun Lai)

指導教授 ： 歐陽明

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202001169

摘要

本論文使用卷積神經網路，將正面及側面兩張二維膝蓋X光影像，重建為三維電腦斷層立體影像。傳統上，取得骨骼三維資訊需要使用電腦斷層掃描。然而，比起X光攝影，電腦斷層掃描的價格較昂貴、輻射量較高、掃描時間較長、且機材可及性較低。本論文與亞東紀念醫院合作，取得歷史病患之X光及電腦斷層影像，並進行資料前處理，得到成對訓練資料。再利用X光攝影方向的特性，建構出能合併兩張輸入之二維影像，輸出三維立體影像的卷積神經網路。
本論文建構之神經網路訓練成果，在能確保X光拍攝角度互為垂直的模擬生成資料集中，成功重建包含骨骼位置、形狀、關節腔凹凸細節之三維影像。而在真實拍攝的X光影像資料集中，雖無法直接輸入神經網路重建出可接受之三維影像，但去除肌肉組織在X光中造成之淡色背景後，可以經由模擬生成資料集所訓練之神經網路重建出骨骼位置、形狀、關節腔輪廓之三維影像。
本論文主要目的並非取代電腦斷層掃描與專業醫師診斷，而是在重建三維資料並視覺化後，能增進醫師與病患溝通其骨骼病徵狀態。

關鍵字

醫療影像 ； 卷積神經網路 ； 膝蓋骨骼

並列摘要

This work conducts a reconstruction of knee bone 3D volume based on frontal and lateral view X-ray images. Traditionally, computed tomography (CT) scan is used to retrieve 3D bone information. However, comparing to X-ray examinations, CT scans are more expensive, incur more radiation dose, require longer examination time, and less accessible. This work acquires historical data of X-ray and CT images from the database of Far Eastern Hospital. After data preprocessing, we obtain paired training data for our neural network. Using the characteristic of the viewing directions of X-ray, we construct a convolutional neural network that can combine two input 2D image, then output a 3D volume.
The resulting convolutional neural network model of this work can successfully reconstruct 3D volume with knee bone position, shape, and cavity surface details, trained by simulated X-ray dataset which ensures that the input images are orthogonal. Although acceptable 3D volume cannot be reconstructed directly in the real X-ray dataset, after removing light background cause by muscle and soft tissues in the X-ray, acceptable 3D volume with knee bone position, shape, and cavity contour can be reconstructed by the neural network model trained by simulated X-ray dataset.
The main purpose of this work is not intent to replace CT examinations or making diagnosis, but to help physician and patient communication by visualizing the reconstructed 3D information of the patient’s knee bone.

並列關鍵字

Medical Imaging ； Convolutional Neural Network ； Knee Bone

參考文獻

Çiçek, Özgün Abdulkadir, Ahmed Lienkamp, Soeren Brox, Thomas Ronneberger, Olaf. (2016). 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation.

Ying, Xingde Guo, Heng Ma, Kai Wu, Jian Weng, Zhengxin Zheng, Yefeng. (2019). X2CT-GAN: Reconstructing CT From Biplanar X-Rays With Generative Adversarial Networks. 10611-10620. 10.1109/CVPR.2019.01087.

Kasten, Yoni Doktofsky, Daniel Kovler, Ilya. (2020). End-To-End Convolutional Neural Network for 3D Reconstruction of Knee Bones From Bi-Planar X-Ray Images.

Ebner, M., Wang, G., Li, W., Aertsen, M., Patel, P. A., Aughwane, R., Melbourne, A., Doel, T., Dymarkowski, S., De Coppi, P., David, A. L., Deprest, J., Ourselin, S., Vercauteren, T. (2019). An automated framework for localization, segmentation and super-resolution reconstruction of fetal brain MRI. NeuroImage, 116324.

Ebner, M., Wang, G., Li, W., Aertsen, M., Patel, P. A., Melbourne, A., Doel, T., David, A. L., Deprest, J., Ourselin, S., Vercauteren, T. (2018). An Automated Localization, Segmentation and Reconstruction Framework for Fetal Brain MRI. In Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2018 (pp. 313–320). Springer.