Automated parameter selection for segmentation of tube-like biological structures using optimization algorithm and mdl

Segmentation of tube-like biological structures, e.g., neurons and blood vessels, allows extraction of structural measurements supporting quantitative studies ranging from understanding neural growth to cancer research. While automated segmentation algorithms minimize subjectivity associated with tedious manual segmentation, they generally have parameter settings to cope with high variability in image data across applications. Currently, these settings are chosen empirically, formulated heuristically, or by trial-and-error with no assurance towards optimality. This work is motivated by the need to automatically select parameter settings for segmentation algorithms since they directly affect segmentation accuracy. An objective trade-off between a probabilistic measure of image-content coverage of a segmentation and its conciseness is based on the minimum description length principle (MDL). The recursive random search (RRS) optimization algorithm is used to efficiently explore combinations of segmentation algorithm parameter settings. For 3-D images, computation time is reduced by coordinated-optimizations on non-empty, representative subimages based on intensity and structural information. The method is initially applied to 223 2-D images of human retinal vasculature and cultured neurons, from four different sources, using a single segmentation algorithm with 8 parameters. Relative to default settings, improvements in the proposed MDL based segmentation quality metric are strongly correlated with improvements in agreement with ground truth (r = 0.78), ranging between 4.7--21% using 1000 function evaluations. Paired t-tests showed that improvements are statistically significant ( p < 0.0005). Most of the improvement occurred in the first 44 function evaluations. For 223 images, RRS outperforms other optimization algorithms (controlled random search, multi-start pattern search, simulated annealing, and genetic algorithm) on average at all 1000 function evaluations. RRS with 1000 function evaluations is on average within 3.56% of the global optimum (6,804,000 function evaluations). The proposed coordinated subimage-optimization method results in average speedup of 11.2X for 22 3-D neurons images. This enables non-expert users to use segmentation algorithms without knowledge of underlying algorithms, increases objectivity, and broadens applicability of the segmentation algorithm. It simplifies the user interface to just one optional parameter, creating a consistent interface that allows developers to freely modify the algorithms. The method allows adaptation of parameters across batches of images and delivers higher morphometric accuracy.