A comparative analysis of deep learning architectures with data augmentation and multichannel input for locoregional...

Studies on deep learning dose prediction increasingly focus on 3D models with multiple input channels and data augmentation, which increases the training time and thus also the environmental burden and hampers the ease of re-training. Here we compare 2D and 3D U-Net models with clinical accepted plans to evaluate the appropriateness of using less computationally heavy models.

A 2D Attention U-Net, a 2D HD U-Net, and a 3D U-Net were trained using 1 or 5 input channels with or without data augmentation and data from 89 locoregional breast cancer patients. Results were compared to clinically accepted plans. The significance of inclusion of more channels or data augmentation was compared to the base models and the HD U-Net and Attention U-Net were compared to their respective identically trained counterparts.

The Attention U-Net reached fewest PTV clinical goals (28%, mostly due to a too high average breast PTV dose) and improved using significantly using five channels and augmentation (49%). The HD U-Net already fulfilled 70% of the PTV goals, which did not improve much by adding more channels or augmentation. The 3D U-Net with five channels and augmentation reached 76%, compared to 81% in the clinically accepted plans. The lower rates for the HD U-Net compared to the 3D U-Net and clinical plans were mainly caused by a lower PTVn1n2 D98%, which was still on average 93%. Organ-at-risk goals were met in most cases for all models. Training time was approximately 8 fold for the 3D model.

Comparable results to a 3D U-Net and clinical plans can be reached with a 2D HD U-net for a dataset size commonly seen in clinical practice. The Attention U-Net did profit from adding extra channels and data augmentation but did not reach the same level of accuracy as the other models.