Multiple Instance Cancer Detection by Boosting Regularised Trees
We propose a novel multiple instance learning algorithm for cancer detection in histopathology images. With images labelled at image-level, we first search a set of region-level prototypes by solving a submodular set cover problem. Regularised regression
- PDF / 2,249,893 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 99 Downloads / 229 Views
1
Introduction
Multiple instance learning (MIL) has recently been applied to histopathology image analysis for both segmentation and classification tasks [10,16]. While training a tumour detector usually requires a large amount of high quality manual annotation [1], MIL methods can potentially infer tumour regions with image-level annotation, i.e., binary labels indicating whether tumour is present in the image. The MIL formulation is attractive as it does not require the effort of manually delineating image regions. The general MIL inference rules are defined in the context of binary classification: a bag of instances is positive if at least one instance in the bag is positive, negative if all of the instances in the bag are negative. A common implementation of the rules in image classification treats each image as a bag, and regions in an image as instances. In terms of histopathology image analysis, an example application is to label an image as cancer if cancer is present in at least one region of the image, and as non-cancer otherwise. In this paper, following the MIL setting, we propose a novel tree boosting algorithm for training a cancer detector. Our algorithm extends Multiple Instance Boosting (MILBoosting) [17] by boosting regularised trees with instanceto-prototype distances as features. The discriminative prototypes in our algorithm are searched by solving a submodular set cover problem. Our approach is validated on two types of histopathology images, namely, breast cancer tissue microarray (TMA) images and optical projection tomographic (OPT) images of colorectal polyps.
2
Related Work
Although MIL has been extensively studied since [12] and there exists a large literature (for a general review of MIL, see [2]), it was only recently applied to c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part I, LNCS 9349, pp. 645–652, 2015. DOI: 10.1007/978-3-319-24553-9_79
646
W. Li, J. Zhang, and S.J. McKenna
histopathology image analysis. Here we give a brief review of some of the most relevant work. Zhao et al. [19] applied MILES [4] for 10 category histopathology image classification. Xu et al. [16] extended MILBoosting [17] to simultaneously detect and cluster multiple types of tissue region in TMA images. Kandemir et al. [9] evaluated MIL formulations on diagnosis of Barrett’s cancer with H&E images. Xu et al. [15] used MIL to classify colon cancer histopathology images with features extracted from convolutional neural networks. Selecting instances as prototypes for bag classification was used previously with bags represented in terms of distances to prototypes [4,7]. Our work extends MILBoosting to select prototypes with instance-to-prototype distances. We search a set of positive instance prototypes that is both discriminative and covers multiple modes of the appearance distribution. Instance-to-prototype distances are considered as features. A regularised regression tree boosting method is proposed to further select and combine the features. Prototypes should
Data Loading...
