Automatic detection and recognition of lung cancer during mass screening of spiral computer tomographic (CT) chest scans is one of the most important problems of today's medical image analysis. We propose an algorithm for isolating lung abnormalities (nodules) from arteries, veins, bronchi, and bronchioles after all these objects have been already separated from the surrounding anatomical structures. The separation is presented elsewhere, and this paper focuses on nodule detection using deformable 3D and 2D templates describing typical geometry and gray level distribution within the nodules of the same type. The detection combines normalized cross-correlation template matching by genetic optimization and Bayesian post-classification. Experiments with 200 spiral low dose CT (LDCT) scans confirm the accuracy of our approach.