Improved Shape Particle Filters for Radiological Image Segmentation

Lukas Fischer, Rene Donner, Franz Kainberger, Georg Langs

Purpose/Introduction: In recent years segmentation approaches based on sequential Monte Carlo Methods delivered promising results for the localization and delineation of anatomical structures in medical images. Also known as Shape Particle Filters, they were used for the segmentation of human vertebrae, lungs and hearts, being especially well suited to cope with the high levels of noise encountered in MR data and difficult overlaps in radiographs. They require a region template of the appearance features which allow to estimate the confidence in the hypotheses generated during the search. Different approaches on 
creating such templates (manual, automatic) exist, but all unnecessarily incorporate the whole template area and even an area around the shape for the actual segmentation. We propose a Differential Evolution Monte Carlo (DE-MC) segmentation scheme using a novel feature extraction method called the Monogenic Signal with considerable speed-up and equal or even better segmentation accuracy compared to the existing approaches. 

Materials and Methods: Objects in images are represented using statistical models of the objects’ shape, defined by landmarks of points allocated at the shape's contour. The modes of shape variation are computed on a set of training images. The pixel coordinates inside the resulting mean shape are stored, forming a per-pixel region map. Calculating the distance between a newly generated shape and the predefined region map yields a residual allowing the Differential Evolution to robustly converge.

Results: The method was evaluated using a) automatic and b) per-pixel region maps on three annotated data sets: 1) hand radiographs, 2) heart MRI slices. The mean distance between segmentation results and ground truth was: 1) a) 7.21, b) 6.7; 2) a) 5.10, b) 4.2. The speed gain was: 1) up to 16 x; 2) up to 15 x faster than 2).

Discussion/Conclusion: The achieved acceleration of the accurate and robust particle filters finally allows to employ them in clinical practice or use them for large scale studies. Even more critically, the extension to 3D is straight forward, extending the advances to further problem domains.