Level set image segmentation with velocity term learned from data with applications to lung nodule segmentation

Abstract: Purpose: Lung nodule segmentation, i.e., the algorithmic delineation of the lung nodule surface, is a fundamental component of computational nodule analysis pipelines. We propose a new method for segmentation that is a machine learning based extension of current approaches, using labeled image examples to improve its accuracy. Approach: We introduce an extension of the standard level set image segmentation method where the velocity function is learned from data via machine learning regression methods, rather than a priori designed. Instead, the method employs a set of features to learn a velocity function that guides the level set evolution from initialization. Results: We apply the method to image volumes of lung nodules from CT scans in the publicly available LIDC dataset, obtaining an average intersection over union score of 0.7185($\pm$0.1114), which is competitive with other methods. We analyze segmentation performance by anatomical and appearance-based categories of the nodules, finding that the method performs better for isolated nodules with well-defined margins. We find that the segmentation performance for nodules in more complex surroundings and having more complex CT appearance is improved with the addition of combined global-local features. Conclusions: The level set machine learning segmentation approach proposed herein is competitive with current methods. It provides accurate lung nodule segmentation results in a variety of anatomical contexts.