ESTRO 2024 - Abstract Book

S4961

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

This study chose to compare the performance of DenseNet with three state-of-art models developed on the HECKTOR 2022 dataset (architectures depicted in Figure 1). TransRP [3] is a DenseNet121-ViT (Vision Transformer) combined model for learning both local detailed and global image features, with early fusion of CT and PET. XSurv [4] performed multi-task learning for GTV segmentation and RFS prediction, with Swin Vision Transformer (SViT) [5] based cross attention to do fusion of CT and PET features at different levels. AdaMSS [6] finetuned the pretrained GTV segmentation network for RFS prediction. It used an attention-based adaptive fusion CNN to fuse CT and PET features at different levels. Their performances were compared to 3D DenseNet with different numbers of layers, early fusion or late fusion, and with or without the GTV contours. Early fusion and late fusion fuse the CT and PET at the input level and the output level of models, respectively. The TransRP, XSurv and AdaMSS were trained using their official codes and data preprocessing. The DenseNet models were trained using the official code of TransRP except replacing the TransRP model to the default DenseNet settings in MONAI 1.1.0 package. The final predictions in the test sets were the averaged ensemble of predictions from the five fold models.

Results:

For the state-of-art models, TransRP and XSurv achieved highest C-indexes of 0.69 and 0.62 in the internal and external test sets, respectively. When inputting CT, PET and GTV based on the early fusion, the optimal number of layers for the DenseNet appeared to be 81. DenseNet81obtained the highest internal test C-index of 0.69 which was similar to that of the TransRP network. However, the performance in the external test set was much lower (C-index