Author: Xiance Jin 👨🔬
Affiliation: 1st Affiliated Hospital of Wenzhou Medical University 🌍
Purpose:
Deep learning deformable registration models was proposed to predict optimal dose distributions a with a few of optimal planned doses using a few-shot learning for cervical cancer.
Methods: A total of 200 VMAT and 200 IMRT plan datasets were utilized with six VMAT plans and 10 IMRT plans selected as test cases and replanned to obtain optimal doses. A deformable dose-structure model, DeformGan, was trained to register doses and anatomies between patients. Optimized plans from other patients were registered to the anatomy of the selected patient and averaged to generate the predicted dose distribution.
Results:
For IMRT, while maintaining the dose coverage on PTV, the DeformGan achieved prediction doses closer to the optimal one on various organs with decreased dose irradiation on the bladder, rectum, and femoral heads. The dose coefficients for unoptimized, optimized, and predicted optimized doses were0.35/0.25/0.27, 0.42/0.20/0.16, 0.67/0.49/0.54, and 0.65/0.43/0.46 on Bladder V40Gy, Rectum V40G, Left femoral head V30Gy, and Right femoral head V30G, respectively. For VMAT, the DeformGan achieved prediction doses closer to the optimal plan in the bladder and rectum in 5 ouf of 6 test patients, while maintaining the dose coverage on PTV and PGTVs.
Conclusion: DoseMorph demonstrated strong performance in predicting optimal dose distributions for IMRT and VMAT in cervical cancer. The model requires minimal optimization plan data during plan optimization and has potential to improve the plan quality efficiently.