Abstract
1745
Objectives Target volume delineation in positron emission tomography remains a challenging task owing to its low spatial resolution and high noise characteristics. As a result a wide variety of image segmentation approaches have been proposed. The objective of this work is to present a novel image segmentation approach based on ant colony optimization algorithms (ACO).
Methods The ACO approach is inspired by the observation of the behavior of real ant colonies based on their collective foraging behavior where ants use pheromone trails as a tool of communication. At the beginning of the ACO algorithm, all the image voxels are initialized with the same pheromone quantity, while at each ant step a certain pheromone quantity is laid by the ants depending on the degree of similarity between the food source and the occupied voxel. The ants start exploring the image looking for pixels that are similar to the food, updating the food source by laying pheromones within pixels. Subsequently a simple analysis of the pheromone distribution leads to image segmentation. 4 patient datasets as well as 13 real simulated tumors were used in the validation and preliminary evaluation of the developed approach in comparison to current state of the art in PET image segmentation using the Fuzzy Locally Adaptive Bayesian approach and manual expert segmentation.
Results Results on simulated datasets show that ACO and FLAB lead to segmentation maps with classification errors of 4.9 ± 3% (median 4.2%, range 1.7 - 11.1%) and 7.4 ± 6.2% (median 5.41%, range 1.4 - 22.5%) respectively..Concerning the 4 patient datasets, the ACO algorithm led to lower classification errors compared to FLAB by considering the manual segmentation by three experts(7.4 ± 6.2% and 17.25 ± 3.99% respectively).
Conclusions An image segmentation approach based on ant colony optimization algorithms was developed and validated. Small errors were obtained in comparison to the ground-truth. Future studies will be based on extending the ACO model to work on 3 tumor classes necessary for a better delineation of heterogeneous tumors.