Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Deep Learning for MRI-Generated Synthetic CT: Dosimetric Evaluation for RT Planning in Head and Neck Cancers

JT Antunes1*, D Pittock1, P Jacobs1, AS Nelson1, J Piper1, T Young2,3, S Deshpande3,4 (1) MIM Software Inc, Cleveland, OH, (2) Institute of Medical Physics, School of Physics, University of Sydney, Australia, (3) Liverpool And Macarthur Hospital, Liverpool, Australia, (4) South Western Sydney Clinical School, University of New South Wales, Sydney, Australia

Presentations

PO-GePV-M-127 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Modern radiation therapy (RT) utilizes both computed tomography (CT) and magnetic resonance imaging (MRI) for treatment planning purposes– CT providing electron density values and MRI providing superior soft-tissue contrast. MR-only guided RT would simplify the clinical workflow and reduce patient burden, but requires methods to derive electron density information from MR images. We present in this work a deep learning-based synthetic CT (sCT) generation framework using a single pre-contrast T1 MRI sequence, and evaluate the dose calculation accuracy in the context of head and neck cancers.

Methods: For 54 patients treated for head and neck cancer, a planning CT (pCT) was acquired the same day as a T1 Dixon MRI scan using the same patient positioning and immobilization. 46 registered MR-CT pairs were used to train a Unet-based convolutional neural network (CNN) to generate sCTs for dose calculation. The CNN included residual blocks and skip connections that transfer high resolution features. The network was trained iteratively to improve the quality of the sCT generated by reducing pixel-wise error with the pCT. The model was validated on 8 unseen MR-CT pairs, with the clinical pCT treatment plan recalculated on the sCT and compared using gamma analysis and dose volume histogram (DVH) comparison.

Results: Our CNN achieved a pixel-wise difference of 78.2±32.5 HU between pCTs and sCTs. DVHs calculated on sCTs were comparable (0.0±1.2 Gy difference) to those calculated on pCTs (Table 1), along with 3D Gamma Index pass rates (96.4-99.5%). Figure 1 displays qualitative comparisons of the sCT and pCT for one test subject, and the corresponding isodose maps and DVH plots.

Conclusion: The comparable calculated dose demonstrates the efficacy of sCT images in treatment planning. Adopting the proposed method offers the potential to eliminate unnecessary pCT scans and may enable MR-only guided RT.

Funding Support, Disclosures, and Conflict of Interest: JA, DP, PJ, AN, and JP are all employees of MIM Software Inc and were reimbursed for travel expenses.

Keywords

Radiation Therapy, Image Processing, Dose Volume Histograms

Taxonomy

IM/TH- MRI in Radiation Therapy: Synthetic CT

Contact Email

Share: