Non-rigid image alignment baed on a local deformation of kernel density functions

2016 Fiscal Year Final Research Report

• Project/Area Number: 15K16087
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Life / Health / Medical informatics
• Research Institution: Kyushu Institute of Technology
• Principal Investigator: Terumasa Tokunaga 九州工業大学, 大学院情報工学研究院, 准教授 (50614806)
• Project Period (FY): 2015-04-01 – 2017-03-31
• Keywords: 統計学 / ベイズ推論 / 機械学習 / 画像処理 / データ科学 / 計算統計

Outline of Final Research Achievements

Recent rapid advance of imaging techniques has produced a massive amount of imaging data in many fields, including biology and medical imaging. Image alignment ( or image registration ) is a process of transforming different set of imaging data measured at different environment into a common coordinate system. It is a core tool for analyzing fMRI, CT, MR images and other medical images. However, the existing alignment techniques often fail to aligning tubular objects ( e.g., blood vessels and neurites). In this work, we developed a new non-rigid image alignment technique based on the idea of a local deformation of kernel density functions. The proposed method begins by performing kernel density estimation to convert digital images into a continuous function. Then, the alignment problem is solved by minimizing the distance between two probability density functions based on EM algorithm. In some experiments, our method achieved a robust optimization for aligning tubular objects.

Free Research Field

データサイエンス、統計的機械学習、計算統計