Synthetic Computed Tomography generation using deep-learning for female pelvic radiotherapy planning

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Synthetic Computed Tomography (sCT) is crucial for electron density mapping in MR-only radiotherapy.
• Current sCT generation methods have limitations in dose accuracy.
• Deep learning (DL) offers a promising approach for improving sCT quality.

Purpose of the Study:

• To develop and evaluate a DL-based method for generating accurate sCTs from T2 spin echo sequences.
• To assess the dosimetric accuracy of DL-generated sCTs compared to traditional methods.
• To determine the feasibility of using routine MRI sequences for sCT generation.

Main Methods:

• A cycleGAN-inspired deep learning model was trained using 30 female pelvis MRI datasets.
• The model generated sCTs from T2 spin echo sequences.
• Dosimetric evaluation involved comparing dose distributions between synthetic CT (sCT) and deformed planning CT (dCT) using mean dose differences and 3D Gamma analysis.

Main Results:

• The DL model achieved a mean dose difference of 0.2% (D98 %) between sCT and dCT.
• Three Dimensional Gamma analysis showed a mean agreement of 90.4% at 1%/1mm.
• Accurate sCTs were generated from standard T2 spin echo sequences.

Conclusions:

• Deep learning enables the generation of accurate synthetic CT images for MR-only radiotherapy.
• The proposed method provides dosimetrically accurate sCTs using readily available MRI sequences.
• This approach eliminates the need for specialized MRI sequences, streamlining radiotherapy workflow.