Technical reports / Department Informatik
- A Cost Constrained Boosting Algorithm for Fast Object Detection (2013)
- Boosting methods are among the most widely used machine learning techniques in practice for various reasons. In many scenarios, however, their use is prevented by runtime constraints. In this paper we propose a novel technique for reducing the computational complexity of hierarchical classifiers based on AdaBoost, such as the probabilistic boosting tree, which are often used for object detection. We modify AdaBoost training so that the hypothesis generation is no longer based solely on the weak learner’s training error but also on a measure of hypothesis complexity. This is achieved by incorporating a cost function into the optimization process, effectively constraining feature selection, which leads to a reduced overall classifier complexity and thus shorter evaluation times. The validity of the approach is shown in an experimental valuation on real-world data. In a cross validation experiment with a system for automatic segmentation of liver tumors in CT images, the evaluation cost for classifying previously unseen samples could be reduced by up to 76% using the methods described here without losing classification accuracy.
- Learning a Prior Model for Automatic Liver Lesion Segmentation in Follow-up CT Images (2013)
- Liver tumors that are not surgically removed need to be closely monitored. A common procedure for their assessment involves acquiring CT images every few months and rating disease status based on the largest diameters of a subset of the lesions. The most prominent benefits of automatic lesion segmentation methods in this context are minimization of time consuming interaction and the possibility of volumetric measurements. While existing methods could be applied to each image individually, we propose to incorporate information gained from previous images of the same patient to enhance the segmentation. We learn a Probabilistic Boosting Tree that has an internal representation of tumor growth from a set of training images. Provided a baseline lesion segmentation, it can generate a patient specific lesion prior to guide the segmentation in a follow-up image. In this paper, we describe and compare different methods for building the growth model and integrating it into a segmentation system. The validity of the approach is shown in an experimental evaluation on a database of 14 patients. On the 17 pairs of baseline and follow-up images in this database, segmentation performance was measured once without and once with the proposed prior. When comparing the points of 90% sensitivity from each experiment, introducing the prior improved the precision of the segmentation from 82.7% to 91.9%. This corresponds to a reduction of the number of false positive voxels per true positive voxel by 57.8%.