Learning to Segment: Training Hierarchical Segmentation under a Topological Loss
We propose a generic and efficient learning framework that is applicable to segment images in which individual objects are mainly discernible by boundary cues. Our approach starts by first hierarchically clustering the image and then explaining the image
- PDF / 1,998,703 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 71 Downloads / 193 Views
2
Institute of Neuroinformatics UZH/ETHZ, Switzerland CellNetworks and IWR/HCI, Heidelberg University, Germany 3 Universitat Pompeu Fabra, Spain
Abstract. We propose a generic and efficient learning framework that is applicable to segment images in which individual objects are mainly discernible by boundary cues. Our approach starts by first hierarchically clustering the image and then explaining the image in terms of a costminimal subset of non-overlapping segments. The cost of a segmentation is defined as a weighted sum of features of the selected candidates. This formulation allows us to take into account an extensible set of arbitrary features. The maximally discriminative linear combination of features is learned from training data using a margin-rescaled structured SVM. At the core of our formulation is a novel and simple topology-based structured loss which is a combination of counts and geodesic distance of topological errors (splits, merges, false positives and false negatives) relative to the training set. We demonstrate the generality and accuracy of our approach on three challenging 2D cell segmentation problems, where we improve accuracy compared to the current state of the art.
1
Introduction
Accurately segmenting a large number of objects of similar type in crowded images is challenging, e.g. when cells of interest touch or overlap. Therefore, the development of robust and efficient algorithms, generic enough to be applicable for different scenarios, is of great importance. Since local measurements are usually ambiguous due to noise and imaging artefacts, priors about the objects to be segmented are needed. Two of the main challenges for automatic segmentation are how to reflect those priors in the cost function of the segmentation problem, and how to learn them from training data. Usually, the cost functions are designed “by hand” for a particular segmentation problem, or have user-adjustable tuning parameters like a foreground bias to adapt to different setups [15,14]. In this paper, we propose a generic framework for structured learning of the cost function for cell segmentation. The cost function is defined on candidate segments, of which we find a cost-minimal, non-overlapping subset to obtain a final segmentation. The main contributions of our approach are: 1) The novel counting-and-propagating topological loss is simple and generic. 2) Our formulation supports a large set of expressive features on the segment candidates. 3) Optimal feature weights are learned from annotated samples to minimize a c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part III, LNCS 9351, pp. 268–275, 2015. DOI: 10.1007/978-3-319-24574-4_32
Learning to Segment
269
topological loss on the final segmentation. The capacity to combine and weigh automatically all features from a large set can reduce the effort previously required to manually select suitable features and tune parameters. 4) By considering candidate segments obtained by iteratively merging superpixels [3,4,7], our method
Data Loading...
