We propose in this paper a novel approach to the automatic segmentation of lung nodules in a given volume of interest (VOI) from high resolution multi-slice CT images by dynamically initializing and adjusting a 3D template and analyzing its cross correlation with the structure of interest. First, thresholding techniques are used to separate the background voxels. The structure of interest, comprising of a nodule candidate and possible attached vessels, is then extracted by excluding any part of the chest wall inside the VOI. Afterwards, the proposed segmentation method finds the core of the structure of interest, which corresponds to the nodule, analyzes its orientation and size, and initializes a 3D template accordingly. Next, The template gradually expands, with its cross correlation to the original structure of interest being computed at each step. The template is then optimized based on the analysis of the cross correlation curve. A segmentation of the nodule is first roughly obtained by doing an 'AND' operation between the optimal template and the extracted structure and then refined by a spatial reasoning method. Template parameters can be recorded and recalled in later diagnosis so that reproducibility and consistency can be achieved. Preliminary results show that segmentation results are consistent, with a mean intra-scan volume measurement deviation of 2.8% for phantom data and 8.1% for real patient data.