Exhibit Hall | Forum 3
Purpose: We propose a sequential transduction neural network with patient-specific self-supervised domain adaptation to carry out spatio-temporal tumor segmentation on longitudinal CBCTs and incorporate the model into a feed-forward adaptive treatment planning strategy.
Methods: We develop a model with a state-of-the-art U-Net for spatial feature extraction and two types of additional sequence transduction architecture including convolutional long short-term memory (LSTM) and transformer to address the temporal features extracted from a series of on-board CBCTs. To overcome the lack of labeled data as well as the degradation of image quality, we leverage a self-supervised domain adaptation technique using a spatial encoder-decoder model jointly trained with a large dataset with pairs of labeled CT and unlabeled CBCTs. In terms of uncertainty estimation, we apply a hybrid of MC dropout and deep ensemble and quantify the uncertainty as the entropy of predictive posterior distributions.
Results: Our experimental results based on a clinical non-small cell lung cancer (NSCLC) dataset with sixteen patients show that our model properly learns weekly deformation of the tumor over time with an average Dice score of 0.92 on the immediate next step, and is able to predict multiple steps (up to 5 weeks) in the future during the patient's treatment with an average Dice score reduction of 0.05, which shows significant improvement over the state-of-the-art U-Net based on direct segmentation of CBCT images, as well as the deformable image registration algorithm. Our proposed method demonstrates a significant reduction in the mean lung dose up to 23% while maintaining the high tumor coverage.
Conclusion: We propose a model based on sequential self-supervised domain adaptation which transfers the knowledge of high-quality pre-treatment CT images to the series of mid-treatment CBCT images. Our results show that our proposed model can successfully predict tumor’s shape which can be used to reduce lung dose.