Segmentation with Incremental Classifiers Guillaume Bernard 1,2 , Michel Verleysen 2 , and John A. Lee 1,2⋆ 1 Molecular Imaging, Radiotherapy, and Oncology – IREC 2 Machine Learning Group – ICTEAM Universit´ e catholique de Louvain, Belgium Abstract. Radiotherapy treatment planning requires physicians to de- lineate the target volumes and organs at risk on 3D images of the patient. This segmentation task consumes a lot of time and can be partly auto- mated with atlases (reference images segmented by experts). To segment any new image, the atlas is non-rigidly registered and the organ contours are then transferred. In practice, this approach suffers from the current limitations of non-rigid registration. We propose an alternative approach to extract and encode the physician’s expertise. It relies on a specific clas- sification method that incrementally extracts information from groups of pixels in the images. The incremental nature of the process allows us to extract features that depend on partial classification results but also convey richer information. This paper is a first investigation of such an incremental scheme, illustrated with experiments on artificial images. 1 Introduction Cancer treatment with radiation beams amounts to a ballistic problem where the dose to the tumor must be maximized while the dose at surrounding healthy tissues must be minimized to avoid secondary effects. In order to achieve the best tradeoff, 3D images of the patients must be segmented to identify the tu- mor and the organs at risk. The physicians use an electronic pen or a mouse to delineate these volumes on each slice. Although it consumes a lot of time, delineation usually remains manual because it involves complex expertise. This explains why usual image segmentation methods such as histogram thresholding [1], pixel or patch clustering [2], gradient peak detection with active contours [3], or watersheds [4, 5] cannot solve the problem. Many of these methods are unsupervised, even though some of them can take into account some a priori in- formation, such as the expected region shape, size, and edge smoothness. On the other hand, supervised segmentation remains difficult to apply, mainly because the encoding of expertise and a priori information is far from being trivial. The most successful approach is the use of atlases, which are (banks of) images that are segmented beforehand by experts. Atlases can be deformed to match any new image with a non-rigid registration algorithm [6]. Once the two images are aligned, the contours or regions can be propagated from the atlas to the new im- age. This approach suffers from the shortcomings of the registration algorithms ⋆ J.A.L. is a Research Associate with the Belgian F.R.S.-FNRS.