Label Stability in Multiple Instance Learning
We address the problem of instance label stability in multiple instance learning (MIL) classifiers. These classifiers are trained only on globally annotated images (bags), but often can provide fine-grained annotations for image pixels or patches (instanc
- PDF / 382,919 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 51 Downloads / 246 Views
Pattern Recognition Laboratory, Delft University of Technology, The Netherlands 2 The Image Section, University of Copenhagen, Copenhagen, Denmark Biomedical Imaging Group Rotterdam, Erasmus MC, Rotterdam, The Netherlands
Abstract. We address the problem of instance label stability in multiple instance learning (MIL) classifiers. These classifiers are trained only on globally annotated images (bags), but often can provide fine-grained annotations for image pixels or patches (instances). This is interesting for computer aided diagnosis (CAD) and other medical image analysis tasks for which only a coarse labeling is provided. Unfortunately, the instance labels may be unstable. This means that a slight change in training data could potentially lead to abnormalities being detected in different parts of the image, which is undesirable from a CAD point of view. Despite MIL gaining popularity in the CAD literature, this issue has not yet been addressed. We investigate the stability of instance labels provided by several MIL classifiers on 5 different datasets, of which 3 are medical image datasets (breast histopathology, diabetic retinopathy and computed tomography lung images). We propose an unsupervised measure to evaluate instance stability, and demonstrate that a performance-stability trade-off can be made when comparing MIL classifiers.
1
Introduction
Obtaining ground-truth annotations for patches, which can be used to train supervised classifiers for localization of abnormalities in medical images can be very costly and time-consuming. This hinders the use of supervised classifiers for this task. Fortunately, global labels for whole images, such as the overall condition of the patient, are available more readily. Multiple instance learning (MIL) is an extension of supervised learning which can train classifiers using such weakly labeled data. For example, a classifier trained on images (bags), where each bag is labeled as healthy or abnormal and consists of unlabeled image patches (instances), would be able to label patches of a novel image as healthy or abnormal. MIL is becoming more and more popular in CAD [9,13,6,20,16,3,21,18,12]. In many of these applications, it is desirable to obtain instance labels, and to inspect the instances which are deemed positive. For example, in [13], weakly labeled x-ray images of healthy subjects and patients affected by tuberculosis are used to train a MIL classifier which can provide local abnormality scores, which can be visualized across the lungs. Furthermore, the MIL classifier outperforms c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part I, LNCS 9349, pp. 539–546, 2015. DOI: 10.1007/978-3-319-24553-9_66
540
V. Cheplygina et al.
its supervised counterpart which has access to fine-grained labels, showing the potential of MIL for CAD applications. A pitfall in using MIL classifiers to obtain instance labels is that these labels might be unstable, for example, if a different subset of the data is used for training. This is clearly undesirable
Data Loading...
