Click here to

  1. Home
  2. Meeting Program
  3. All Sessions
  4. Abstracts
Session: [Return to Session]

Genetic Algorithm and Neural Networks in Radiosurgery for Multiple Metastases

JA Rojas-Lopez1*, CD Venencia2, MA Chesta1, F Tamarit1, (1) Universidad Nacional de Cordoba, Cordoba, ,AR, (2) Instituto Zunino, Cordoba, X, AR

Presentations

PO-GePV-T-404 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To evaluate the optimization of PTV margins in multiple metastases radiosurgery (SRS) with single isocenter technique by the use of bio-inspired algorithms and neural networks.

Methods: 10 plans were created and optimized with Elements Multiple Mets SRS v2.0 (Brainlab AG, Munchen, Germany). The mean number of metastases per plan was 5±2 [3,9] and the mean volume of GTV was 1.1±1.3 cc [0.02, 5.1]. The total number of metastases was 55. Considering all possible combinations of rotational and translational movements (6!x2⁶₌46080), the maximum displacement (roll, pitch, yaw, x, y, z) was optimized by a genetic algorithm (GA). By the use of a multilayer perceptron, the PTV margin (2 mm, 1 mm or 0.5 mm) was determined considering the target distance to isocenter and the volume of the lesion. The original plans were re-calculated using the PTV optimized margin and new dosimetric variations were obtained. The Paddick conformity index (PCI) and gradient index (GI) were analyzed.

Results: The GA parameters such as number of parents, cross-over point and mutation rate were optimized to reduce the computation time and to obtain global optimization points. Considering the maximum effective displacements due to rotations and translations, it is necessary to define larger and optimized PTV margins to reduce dosimetric variations on PCI and GI. The multilayer perceptron neural networks hyperparameters (learning rate, activation function, inner layers, number of neurons) were optimized for reducing the computation time and to obtain better loss functions.

Conclusion: The GA and neural networks are tools to facilitate the PTV margin decision on SRS for multiple metastases with single isocenter. These computational tools based on artificial intelligence consider a complete dosimetrical and geometrical study of the mechanical uncertainties due to rotations and translations in these treatments.

Keywords

Genetic Algorithm, Radiosurgery, Software

Taxonomy

TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email

Share:

Additional Links