@article{GOMEZ2020106793,
title = "Performance analysis of real-coded evolutionary algorithms under a computationally expensive optimization scenario: 3D–2D Comparative Radiography",
journal = "Applied Soft Computing",
volume = "97",
pages = "106793",
year = "2020",
issn = "1568-4946",
doi = "https://doi.org/10.1016/j.asoc.2020.106793",
url = "http://www.sciencedirect.com/science/article/pii/S1568494620307316",
author = "Oscar Gómez and Oscar Ibáñez and Andrea Valsecchi and Enrique Bermejo and Daniel Molina and Oscar Cordón",
keywords = "Comparative radiography, Computer vision, Evolutionary computation, 3D–2D evolutionary image registration",
abstract = "Real-coded evolutionary algorithms have solved numerous real-world optimization problems. In this work, we aim to analyze the behavior and robustness of several real-coded evolutionary algorithms from the state of the art in a challenging real world optimization problem. This optimization problem consists on the superimposition of 3D and 2D images of skeletal structures (i.e. bones and cavities) based on their silhouette. This task is required for the automation of a forensic identification technique known as comparative radiography, via the generation of the best projection of the 3D image with respect to the 2D image. This superimposition problem was tackled in a recent proposal using an evolutionary 3D–2D image registration method based on differential evolution. However, the results obtained were insufficient for its use in real scenarios, due to: (1) the complexity and multi-modality of search space, despite the reduced number of parameters to be optimized (7 in its simple version and 9 in a more complex one, proposed in this work); and (2) the high computational cost of generating and evaluating a superimposition. Particularly, we have performed a rigorous comparative study of six state-of-the-art real-coded evolutionary algorithms (DE, L-SHADE, CMA-ES, BIPOP-CMA-ES, CRO-SL, and MVMO-SH) with synthetic images of three forensic anatomical structures (frontal sinuses, clavicles, and patellae), showing that the best results are always obtained by MVMO-SH in terms of precision, robustness and computational cost. Furthermore, we have validated the quality of the superimpositions obtained by the evolutionary image registration method using MVMO-SH with real images of frontal sinuses. We have performed the comparison of 50 head radiographs and 50 3D images, resulting in 2,500 cross-comparisons (50 positive and 2,450 negatives). The obtained results are promising since the superimpositions obtained allowed us to filter out 88% of the possible candidates with 0 error rate in a fully automatic manner, showing the high quality of the superimposition obtained."
}