Empowering Vision Transformer by Network Hyper-Parameter Selection for Whole Pelvis Prostate Planning Target Volume Auto-Segmentation

Abstract

Simple Summary Vision transformers have been recently spread out to enhance segmentation accuracy, becoming an active area of research and development involved in radiotherapy. We found that the new network architecture did not guarantee improvement. Conventional CNN-based networks struggled with being expanded to the auto-segmentation of tumors from normal organs due to local geometric dependence and difficulty in the hyper-parameter selection. As seen in the development and success of nnU-Net, we emphasized the importance of finding suitable hyper-parameters for the vision transformer. We applied our proposed framework based on VT U-Net v.2 to the prostate target volume segmentation, followed by extensively validating its performance in segmentation accuracy against the other five competing deep neural networks through four-fold cross-validation using CT images.Abstract U-Net, based on a deep convolutional network (CNN), has been clinically used to auto-segment normal organs, while still being limited to the planning target volume (PTV) segmentation. This work aims to address the problems in two aspects: 1) apply one of the newest network architectures such as vision transformers other than the CNN-based networks, and 2) find an appropriate combination of network hyper-parameters with reference to recently proposed nnU-Net ("no-new-Net"). VT U-Net was adopted for auto-segmenting the whole pelvis prostate PTV as it consisted of fully transformer architecture. The upgraded version (v.2) applied the nnU-Net-like hyper-parameter optimizations, which did not fully cover the transformer-oriented hyper-parameters. Thus, we tried to find a suitable combination of two key hyper-parameters (patch size and embedded dimension) for 140 CT scans throughout 4-fold cross validation. The VT U-Net v.2 with hyper-parameter tuning yielded the highest dice similarity coefficient (DSC) of 82.5 and the lowest 95% Haussdorff distance (HD95) of 3.5 on average among the seven recently proposed deep learning networks. Importantly, the nnU-Net with hyper-parameter optimization achieved competitive performance, although this was based on the convolution layers. The network hyper-parameter tuning was demonstrated to be necessary even for the newly developed architecture of vision transformers.