Assessing Multiple MRI Sequences in Deep Learning-Based Synthetic CT Generation for MR Only-Guided Radiation Therapy of Head and Neck Cancers

Abstract

Listen

Translate article

<h3>Purpose/Objective(s)</h3> MRI is often utilized in radiation therapy (RT) planning for contouring due to its superior soft tissue contrast. However, a planning CT is still required to provide electron density information for dose calculation. Reducing the need for multiple modalities for quality RT planning would improve clinical workflow and reduce patient burden. Synthetic CT generation based on MRI is a promising solution, but it is not yet known which set of MR sequences generates the most accurate synthetic CT. In this work, we assessed 14 combinations of MR sequences as inputs to a deep learning-based synthetic CT generator in the context of MR-guided RT planning for head and neck cancers. <h3>Materials/Methods</h3> 26 patients who underwent RT for head and neck cancer were retrospectively identified and included in this study based on availability of 3 pretreatment MR Dixon sequences (T1pre-contrast, T1post-contrast, T2) and a planning CT. Each Dixon sequence generated an in-, opposed-, fat-only-, and water-only-phase image, yielding 12 MR images per patient. MR and planning CT images were acquired on the same day, with patients scanned in the same treatment position and immobilization. CT images were registered to a primary MR sequence, masked to eliminate background elements, and clipped to a fixed intensity range. Our synthetic CT generator comprised a 2D Unet architecture with skip connections, designed for a variable subset of MR inputs. For each combination of MR sequences, the deep learning model was trained using 21 patients and validated on 5 patients. Dose was computed on both the planning and synthetic CT via Monte Carlo simulation using the clinical RT plan. Performance was measured in terms of mean absolute error (MAE) of HU as well as 3D global gamma (3% / 3mm) between the planning and synthetic CT and their respective doses. <h3>Results</h3> MR sequences yielded average MAE from 78.9 - 86.5 HU. More MR sequences only moderately correlated with lower MAE (R² = 0.613, p > 0.05). T2 sequence had the poorest performance. No differences were observed between synthetic CTs generated using T1pre- and T1post-contrast sequences. All combinations of MR inputs yielded synthetic CTs with dose plans yielding extremely high average gamma pass rates (99.5 - 99.8%). <h3>Conclusion</h3> No significant differences were found among individual MR sequences; the acquisition of more MR sequences for synthetic CT generation did not yield additional improvement in terms of clinical efficacy. Using our proposed method offers the potential for clinically viable MR-guided RT without having to acquire additional sequences for synthetic CT generation.