Lung cancer dose distribution prediction based on a dual-branch feature extraction network.

Background: Currently, predicting dose distributions through neural networks can improve the automation level of radiotherapy planning. However, a single neural network often has limitations in its ability to extract features and obtain clinical information.

Purpose: To assist in the development of volumetric modulated arc therapy (VMAT) plans for non-small cell lung cancer (NSCLC) patients, a dual-branch feature extraction neural network is proposed to predict dose distributions.

Methods: This study proposes a dual-branch feature extraction network named CTNet, which consists of a convolutional network and a transformer network in parallel to extract local and global features that are meaningful for dose prediction tasks. A feature fusion module has been developed to reduce the heterogeneity of the two extracted features. To promote the learning of two types of features in the network, weighted mean square error and multiscale structural loss were used. The network was trained on 144 VMAT plans of NSCLC patients. The performance of this network was compared with that of several commonly used networks, and the network performance was evaluated on the basis of the voxel-level mean absolute error (MAE) within the planning target volume (PTV) and organs at risk (OARs), as well as the error in clinical dose‒volume metrics.

Results: The MAE between the predicted dose distribution and the manually planned dose distribution within the PTV was 1.14 Gy, and the D95 error was less than 1 Gy. Compared with the other four commonly used networks, the dose error of the CTNet was the smallest in the PTV and OARs.

Conclusions: The proposed CTNet uses the transformer and convolutional networks to extract global information, such as the relative position of the PTV and OARs, as well as local information, such as shape and size, enabling accurate prediction of the dose distribution for NSCLC patients undergoing VMAT radiotherapy.