Multi-atlas segmentation using manifold learning with deep belief networks

This paper proposes a novel combination of manifold learning with deep belief networks for the detection and segmentation of left ventricle (LV) in 2D — ultrasound (US) images. The main goal is to reduce both training and inference complexities while maintaining the segmentation accuracy of machine learning based methods for non-rigid segmentation methodologies. The manifold learning approach used can be viewed as an atlas-based segmentation. It partitions the data into several patches. Each patch proposes a segmentation of the LV that somehow must be fused. This is accomplished by a deep belief network (DBN) multi-classifier that assigns a weight for each patch LV segmentation. The approach is thus threefold: (i) it does not rely on a single segmentation, (ii) it provides a great reduction in the rigid detection phase that is performed at lower dimensional space comparing with the initial contour space, and (iii) DBN's allows for a training process that can produce robust appearance models without the need of large annotated training sets.

1. Introduction

Two sequential stages are typical in current machine learning methodologies for object segmentation [1], [2]: (i) rigid detection (coarse step) and (ii) non-rigid segmentation (fine step). The first step (rigid detection) is of crucial importance, since it reduces the search running time and training complexities. This paper achieves a complexity reduction of the rigid detection ¹

State-of-the-art rigid detection produces in practice, translation, rotation and scaling of the visual object, (e.g [3]), i.e. R = 5. In this paper, the rigid detection space achieves M < R, where M in the intrinsic dimension of the manifold. See Section 6.1.

by using a manifold learning algorithm. This is an atlas-based segmentation, in sense that the manifold partitions (by soft clustering) the data into several patches using two distinct assumptions: (i)preserving the angular within a patch using a smaller number of points and (ii) the distance (i.e. neighborhood) within these points [4]. Each patch in the learned manifold provides a segmentation proposal. Since multiple patches are obtained, multiple segmentations should be combined. In this paper, a novel strategy to accomplish this is developed. More specifically, a DBN multi-classifier for final segmentation is proposed. This means that a multi-atlas segmentation strategy is followed, i.e. the different segmentations are fused within patches, in which the weights are given by the deep belief network classifiers.