A Comparative Study of Deep Learning Dose Prediction Models for Cervical Cancer Volumetric Modulated Arc Therapy.

Area of Science:

• Medical Physics
• Radiotherapy Oncology
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) is increasingly utilized in radiation oncology for dose prediction.
• A lack of comparative studies on different DL techniques hinders clinical adoption.
• Accurate dose prediction is crucial for optimizing volumetric modulated arc therapy (VMAT) in cervical cancer treatment.

Purpose of the Study:

• To compare the performance of four state-of-the-art deep learning models.
• To evaluate their accuracy in predicting voxel-level dose distributions for cervical cancer VMAT.
• To identify the most effective DL model for this application.

Main Methods:

• Retrospective analysis of 261 cervical cancer patient plans.
• Utilized 3D U-Net and three variants, trained on CT images, PTV, and OARs masks.
• Evaluated models using mean absolute error (MAE), dose difference, dosimetric indices, and Dice similarity coefficients (DSC) on a testing set.

Main Results:

• All DL models showed promising dose prediction accuracy, with maximum MAE within the PTV of 0.83% (UNETR).
• Maximum MAE for organs at risk (OARs) was observed in the left femoral head (6.95%).
• 3D U-Net achieved the minimum MAE (0.94%) for the entire body, outperforming other models.

Conclusions:

• Deep learning models accurately predict voxel-level dose distributions for cervical cancer VMAT.
• While models showed analogous performance, 3D U-Net exhibited superior results for overall body dose prediction.
• These advanced DL models hold significant potential for clinical application in cervical cancer VMAT.