Browse Prior Art Database

Method for Atlas Reduction in Label Fusion-Based Segmentation Disclosure Number: IPCOM000241433D
Publication Date: 2015-Apr-25
Document File: 6 page(s) / 118K

Publishing Venue

The Prior Art Database


Disclosed is a method for atlas reduction in multi-atlas label fusion that exploits similarity of three-dimensional (3D) data by dividing 3D data into two-dimensional (2D) slices. The proposed method improves accuracy for a given number of atlases and reduces computation time compared to multi-atlas label fusion using 3D atlases.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 30% of the total text.

Page 01 of 6

Method for Atlas Reduction in Label Fusion-Based Segmentation

A reduction in the number of atlases is needed to perform multi-atlas label fusion, which is used for image segmentation

and labeling. This approach is specifically applied to cardiac magnetic resonance images (MRI).

Multi-atlas label fusion is one of the most powerful techniques for segmenting biomedical imaging and has been applied to cardiac MRI. Typically, the use of three-dimensional (3D) images requires a 3D registration and label fusion. Both the 3D registration and label fusion are time consuming and prone to errors, which results in inaccurate segmentation results. Because the 3D images are a series of two-dimensional (2D) slices, reducing this problem to a series of 2D registrations and label fusion steps reduces the computational burden. Additionally, neighboring slices around a target slice can be used as additional atlases because the appearance of these slices is similar. This means that for each slice being segmented, there are N times the number of atlases used for label fusion compared to a traditional approach, where N is

the number of neighboring slices.

A method is needed to reduce the number of atlases and the amount of computational time needed to segment cardiac

MRI images.

The novel solution is an augmented multi-atlas label fusion approach that drastically reduces both the number of atlases and computational time needed to achieve state of the art segmentation accuracy in cardiac MRI. The implementation of this work involves five steps:

1. Preprocessing
2. Landmark generation
3. 2D landmark registration
4. 2D label fusion
5. Post-processing

The method begins with the identification five landmarks used for the alignment of cardiac slices. The landmark identification is currently a manual process, but can be automated in the future. The central slice contains three landmarks that are used for an affine registration. Each 3D atlas is divided into 2D slices and the center slice of each atlas is aligned to the center slice of the target image to be segmented. The same affine transformation is then applied to all other 2D slices for a given atlas. This drastically reduces the computational time when compared to a traditional 3D non-rigid registration. For each 2D slice in the target image, the corresponding 2D slice for each atlas is determined. That slice, the two slices above, and the two slices below are used to label the target slice. In this way, each subject


Page 02 of 6

provides five atlases for label fusion.

As a result, the dice similarity coefficient (DSC) is evaluated for the given approach varying the number of atlases used. For example, for a dataset containing 83 training images and 72 testing images the proposed method can achieve a mean DSC of .818 using only 10 atlases with small improvements until a plateau using 40 atlases with a DSC of .825. These DSC scores are comparable to other state of the art methods that require four times th...