Abstract
PURPOSE: Heterogeneity correction is vital in radiation therapy treatment planning to ensure accurate dose delivery. Brain cancer stereotactic treatments, like Gamma Knife radiosurgery (GKRS), often rely on homogeneous water-based calculations despite the potential heterogeneity impact near bony structures. This study aims to develop a method for generating synthetic dose plans incorporating heterogeneity effects without additional computed tomography (CT) scans. METHODS AND MATERIALS: Magnetic resonance imaging and CT images, TMR10-based, and convolution-based dose distributions were used from 100 retrospectively collected and 22 prospectively collected GKRS patients. A conditional Generative Adversarial Network was trained to translate TMR10 into synthetic convolution (sConv) doses. RESULTS: The generated sConv dose demonstrated qualitative and quantitative similarity to the actual convolution (Conv) dose, showcasing better agreement of dose distributions and improved isodose volume similarity with the Conv dose in comparison to the TMR10 dose (γ pass rate; sConv dose, 92.43%; TMR10 dose, 74.18%. Prescription isodose dice; sConv dose, 91.7%; TMR10 dose, 89.7%). Skull-induced scatter and attenuation effects were accurately reflected in the sConv dose, indicating the usefulness of the new dose prediction model as an alternative to the time-consuming convolution dose calculations. CONCLUSIONS: Our deep learning approach offers a feasible solution for heterogeneity-corrected dose planning in GKRS, circumventing additional CT scans and lengthy calculation times. This method's effectiveness in preserving dose distribution characteristics in a heterogeneous medium while only requiring a homogeneous dose plan highlights its utility for including the process in the routine treatment planning workflows. Further refinement and validation with diverse patient cohorts can enhance its applicability and impact in clinical settings.