Big improvement of the generalization ability and robustness in construction of hybrid hierarchical statistical shape model of anatomical structures

2016 Fiscal Year Final Research Report

• Project/Area Number: 26330191
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Nagoya Institute of Technology
• Principal Investigator: Hontani Hidekata 名古屋工業大学, 工学(系)研究科(研究院), 教授 (60282688)
• Project Period (FY): 2014-04-01 – 2017-03-31
• Keywords: 統計モデリング / モデル選択

Outline of Final Research Achievements

In this project, we developed a new method for constructing statistical shape model of anatomical organs robustly by employing not a normal distribution but a q-exponential distribution for the representation of the variety of the shape in order for estimating appropriate values of parameters even when the number of training data is small. A statistical shape model is constructed from a set of training data and it is not easy to collect the training data as many as the degree of the freedom of the shape variety. When the training data set is small and one employs a normal distribution for the representation, the resultant model overfits to the small training data set and the representation ability is degraded. Each layer of many conventional hierarchical models for the organs is mainly represented by a normal distribution. Using the results of this research, one can automatically select the best representation that maximizes the generalization ability.

Free Research Field

医用画像処理