Nested CNN architecture for three-dimensional dose distribution prediction in tomotherapy for prostate cancer

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• Investigating novel deep learning architectures for enhanced accuracy in radiation dose distribution prediction.
• Exploring network layer modifications as an alternative to dimension expansion for complex dose calculations.

Purpose of the Study:

• To evaluate the efficacy of a nested UNet architecture in improving dose distribution prediction accuracy compared to a standard UNet.
• To assess the performance of these models using dosimetry and geometry indices.

Main Methods:

• A cohort of 137 prostate cancer patients treated with tomotherapy was used, with data split into training/validation (80%) and testing (20%) sets.
• Nested UNet and UNet models were trained and evaluated using Mean Absolute Error (MAE) for dose-volume histograms (DVHs) and similarity indices (SSIM, DSC, JSC) for isodose volume (IV) prediction.
• Statistical significance was determined using the two-way Wilcoxon test (p < 0.05).

Main Results:

• The nested UNet architecture demonstrated reduced MAE in DVH indices across critical structures like the planning target volume (PTV), bladder, and rectum.
• The nested UNet achieved a higher mean Structural Similarity Index Measure (SSIM) of 0.94 compared to the UNet's 0.91, indicating superior geometric similarity prediction.
• Specific MAE values for PTV, bladder, and rectum showed significant improvements with the nested UNet.

Conclusions:

• The nested UNet network is a suitable deep learning model for improving the accuracy of dose distribution prediction in radiotherapy.
• This architecture offers a significant advantage over the standard UNet in terms of predictive accuracy and efficiency.
• The findings support the use of nested UNet for more precise radiotherapy planning.