Premium
WE‐E‐213CD‐06: A Locally Adaptive, Intensity‐Based Label Fusion Method for Multi‐ Atlas Auto‐Segmentation
Author(s) -
Han X
Publication year - 2012
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4736162
Subject(s) - atlas (anatomy) , artificial intelligence , segmentation , computer science , image registration , contouring , computer vision , weighting , image warping , pattern recognition (psychology) , centroid , image segmentation , fusion , image (mathematics) , medicine , linguistics , philosophy , computer graphics (images) , radiology , anatomy
Purpose: Atlas‐based auto‐segmentation (ABAS) has emerged as a very useful contouring tool for radiotherapy planning. Higher accuracy of ABAS typically requires the use of multiple atlases, for which the final label fusion step is a key design component. This work presents a novel locally adaptive, intensity‐based label fusion approach for multi‐atlas ABAS, and compares its performance against the commonly used STAPLE method. Methods: The label fusion method derives the final structure label for a novel patient image as a weighted average of several warped atlas label maps, where the atlas warping is achieved through deformable atlas registration. Instead of assigning a constant global weighting factor for each atlas and for each structure, adaptive weights are computed at each image location based on the local correlation coefficients (LCC) computed between the patient image and each warped atlas image. To compensate for registration errors, neighboring atlas labels within a small distance from the center point are also considered in the fusion computation, but only the first k (typically 25) neighbors with the largest LCC are included to get better accuracy. The method was evaluated using ten manually contoured H&N patient images with a leave‐one‐out validation strategy. Performances of the newly proposed method and the classical STAPLE method are compared for 7 structures including the mandible, the parotids (left and right), the sub‐ mandibular glands (left and right), the brainstem, and the spinal cord. Results: The proposed intensity‐based label fusion method significantly outperforms the STAPLE method for all structures considered. The improvement of the mean Dice value ranges from 1.5%for the right parotid to 9% for the right sub‐mandibular gland. Conclusions: The locally adaptive, intensity‐based label fusion provides a superior accuracy compared to the STAPLE method, which helps boost the performance of ABAS methods and make them more usefulness in practice. The author is a current employee of Elekta Inc.