Feasibility study of using CNN-GRU-Dense model for real-time liver tumor tracking during radiotherapy

Abstract

Background

Real-time tumor tracking helps overcome challenges in delivering accurate radiotherapy. Commercial tracking devices use a hybrid external–internal correlation model (ECM) that combines intermittent X-ray imaging of the tumor's internal position with continuous monitoring of external respiratory motion. This approach improves tracking accuracy and treatment effectiveness.

Purpose

This study simulated using a deep learning model (CNN-GRU-Dense model) to track liver tumors in real-time during treatment—without needing ongoing updates. The model's accuracy was tested against several well-known methods, including the hybrid correlation model used in the CyberKnife system, the NG-RC model, and the augmented linear model.

Methods

The CNN-GRU-Dense model comprises convolutional, Gated Recurrent Units (GRU), and dense layers to estimate tumor position in various directions. Initially, input signals are processed through a 1D convolutional layer that employs 64 filters with a kernel size of 3 and ReLU activation to extract spatial features. Next, the extracted features are processed by two stacked GRU layers, each containing 256 units with ReLU activation, enabling the model to capture temporal dependencies. After the GRU layers, the data undergoes refinement through two dense (fully connected) layers, each with 64 units and ReLU activation, ensuring enhanced feature extraction. Finally, the output is passed through a single-unit output layer with linear activation, providing the estimated tumor position. For training the CNN-GRU-Dense model, 26 min of motion patterns (a patient-specific data) are utilized. The proposed model underwent hyperparameter optimization using the RandomSearch approach. This method explored a broad search space, which included the number of filters and kernel size in the 1D Convolutional layer, the number of GRU units, the number of fully connected dense layers, the learning rate, and the loss function. Using a learning rate 0.001, the model was optimized with the Adam optimizer and trained with the mean squared error (MSE) loss function. The training was conducted for 30 epochs with a batch size of 300, aiming to strike a balance between speed and stability during the learning process. Finally, the trained CNN-GRU-Dense model was tested with new external motion data to estimate tumor positions. The model parameters remain unchanged throughout the treatment, requiring no updates. Fifty-seven motion trace datasets from the CyberKnife system were used to evaluate the CNN-GRU-Dense model performance. These traces were grouped into three liver regions: central, lower, and upper.

Results

The CNN-GRU-Dense model demonstrated improved estimation accuracy compared to other ECMs (Wilcoxon signed rank p < 0.05). The 3D radial estimation accuracy (Mean ± standard deviations (SD)) using the CyberKnife system, the NG-RC model, the augmented linear model, and the CNN-GRU-Dense model was 1.42 ± 0.44 mm, 1.23 ± 0.75 mm, 0.71 ± 0.40 mm, and 0.55 ± 0.27 mm, respectively.

Conclusions

The simulation results showed that the CNN-GRU-Dense model outperformed several existing methods, including the augmented linear model used in standard linear accelerators, the NG-RC model, and the constrained fourth-order polynomial equations used in the CyberKnife and Radixact systems. One key advantage of the CNN-GRU-Dense model is that it doesn't need to be updated during treatment, which reduces patients' radiation exposure.