An Improved Multi-task Learning Approach with Applications in Medical Diagnosis Jinbo Bi 1 , Tao Xiong 2 , Shipeng Yu 1 , Murat Dundar 1 , and R. Bharat Rao 1 1 CAD and Knowledge Solutions, Siemens Medical Solutions 20 Valley Stream Parkway, Malvern, PA 19355, USA jinbo.bi@siemens.com 2 Risk Management, Applied Research, eBay Inc. 2145 Hamilton Avenue, San Jose, CA 95125, USA Abstract. We propose a family of multi-task learning algorithms for collaborative computer aided diagnosis which aims to diagnose multiple clinically-related abnormal structures from medical images. Our formula- tions eliminate features irrelevant to all tasks, and identify discriminative features for each of the tasks. A probabilistic model is derived to justify the proposed learning formulations. By equivalence proof, some existing regularization-based methods can also be interpreted by our probabilis- tic model as imposing a Wishart hyperprior. Convergence analysis high- lights the conditions under which the formulations achieve convexity and global convergence. Two real-world medical problems: lung cancer prog- nosis and heart wall motion analysis, are used to validate the proposed algorithms. 1 Introduction Physicians routinely use computer aided diagnosis (CAD) systems in clinical practice [1]. It is well accepted that CAD systems decrease detection and recog- nition errors when used as a second reader [2]. Typically, the goal of a CAD system is to detect potentially abnormal structures in medical images. How- ever, most CAD systems focus on the diagnosis of a single isolated abnormality using images taken only for the specific disease, which neglects a fundamental aspect of physicians diagnostic workflow where they examine not only primary abnormalities but also symptoms of related diseases. For instance, an automated lung cancer CAD system can be built to separately identify solid nodules and ground glass opacities (GGOs). (Patients can have both structures, or GGOs can later become calcified GGOs which become solid or partly-solid nodules.) Radiologic classification of small adenocarcinoma of lung by means of thoracic thin-section CT discriminates between solid nodules and GGOs. Fig. 1 shows two CT slices with a nodule and a GGO respectively. A solid nodule is defined as an area of increased opacification more than 5mm in diameter, which completely obscures underlying vascular markings. A ground- glass opacity (GGO) is defined as an area of a slight homogeneous increase in density, which does not obscure underlying vascular markings [3]. Detecting W. Daelemans et al. (Eds.): ECML PKDD 2008, Part I, LNAI 5211, pp. 117–132, 2008. c  Springer-Verlag Berlin Heidelberg 2008