林對正：全髖關節置換手術前後之X光影像對正

Abstract

本論文提出林對正: 一個電腦視覺的演算法，用於將手術前後的X光影像進行對齊，提供一套系統方便醫師用更有效率的方式比對手術前後的影像差異，取代傳統的手動對齊X光片的方法。這方法能夠在對正影像時，只將特定區域進行對正，解決因對象非剛體而無法線性轉換的問題。因此適合應用在進行全髖關節置換手術過程中比對骨骼位置，讓醫師可以檢查植入物是否安裝正確。 在我們的實驗中，我們以骨盆的X光影像作為實驗對象，每組影像為同個病患在不同時間所拍攝的X光照片。我們實驗了多個不同方法，藉由比對影像之間的相似特徵，計算這些特徵點的位移，便可以將影像進行對齊。 我們透過預先定義的對應關鍵點在影像對齊後的誤差來衡量演算法的表現好壞，這些關鍵點是骨骼系統在解剖學上的重要特徵。實驗的目標是最小化影像對齊後的關鍵點距離，計算這些對應點距離的均方差，作為對我們演算法的評分依據。

In this thesis, we propose LinAlign: a computer vision algorithm for aligning X-ray images before and after surgery, providing a system for surgeons to compare images before and after surgery more efficiently and replace traditional manually aligning radiographs. LinAlign allows only align specific area when aligning images, and solve the problem that linear transformation cannot be performed on non-rigid objects. Therefore, it is suitable for comparing the position of bones during the hip replacement surgery, allowing orthopedic surgeons to make sure that implants have been installed correctly. In our experiment, we took the X-ray images of the pelvis as our experimental data: each set of images contains the X-ray photographs of the same patient taken at different times. We experiment with different methods. By comparing similar features between images and calculating the displacement of these feature points, the images can be aligned. We evaluate the performance of the algorithm by the error of pre-defined landmarks after alignment. These landmarks are anatomically important features of the skeletal system. The goal of our experiment is to minimize the distance of landmarks between image pair. We take the mean square error of these landmark distances as the performance metric to our algorithm.