In this work we propose a machine learning approach to improve shape detection accuracy in medical images with deformable contour models (DCMs). Our DCMs can efficiently recover globally optimal solutions that take into account constraints on shape and appearance in the model fitting criterion; our model can also deal with global scale variations by operating in a multi-scale pyramid. Our main contribution consists in formulating the task of learn-ing the DCM score function as a large-margin structured prediction problem. Our algorithm trains DCMs in an joint manner -all the pa-rameters are learned simultaneously, while we use rich local features for landmark localization. We evaluate our method on lung field, heart, and clavicle seg-mentation tasks using 247 standard posterior-anterior (PA) chest radiographs from the Segmentation in Chest Radiographs (SCR) benchmark. Our learned DCMs systematically outperform the state of the art methods according to a host of validation measures includ-ing the overlap coefficient, mean contour distance and pixel error rate. 1. INTRODUCTION Precisely localizing shapes in medical images is of paramount im-portance in a host of medical image applications, involving organ segmentation, tracking, registration and atlas building. In our work we consider the problem of localizing a set of landmarks that are strung together along a contour; our task is to recover this landmark sequence by exploiting both the local appearance information around the individual landmarks, as well as their ordering constraints. Deformable contour models (DCMs) constitute a main workhorse for detecting such 1D structures in medical images -starting from the seminal works of Snakes [1], Deformable Templates [2] and Active Shape/Appearance Models [3, 4], DCMs have been thriving in problems involving shapes for more than two decades. One of the most desirable properties of DCMs is that they allow to cast tasks such as segmentation or tracking in terms of optimization by incorporating the desirable properties of the envisioned solution in the form of a merit function. One can then optimize this function with off-the-shelf techniques, such as Dynamic Programming (DP) [5], Gradient Descent [4], or more dedicated techniques such as curve evolution with Level Sets [6]. We focus on learning the merit function being optimized so as to improve the shape localization performance of DCMs. Earlier re-search has involved enhancing the geometric terms in DCMs, includ-ing their formulation in intrinsic geometric terms [7], the incorpora-tion of better regularization terms [8] and the introduction of shape priors [9] in curve evolution. Current works in medical imaging es-timate the model parameters in a two-stage training manner, using e.g. maximum likelihood (ML) estimation for the pairwise terms,