3595 - A Study on the Correlation Model between Surface Motion and Internal Target Position Based on 3D nnUNet

Purpose/Objective(s):

Respiratory motion affects the effectiveness of lung radiotherapy. Surface Guided Radiotherapy (SGRT) utilizes non-radiative optical imaging to accurately guide patient positioning and monitor respiration in real-time. However, its capability to monitor target motion within the thoracoabdominal region remains limited. This study aims to establish a correlation model between surface and internal target motion using deep learning techniques, in order to achieve precise real-time prediction of lung tumor location and shape based on surface motion signals.

Materials/Methods:

A dataset consisting of 4DCT images from 119 lung cancer patients across multiple centers was used. Based on extracted internal respiratory signals and respiratory surrogate signals, an external-internal motion correlation model for respiratory motion was established using a 3D nnU-Net deep learning algorithm framework. The prediction accuracy of the model was then evaluated under the condition of limiting and fixing the effective input to the body surface deformation region. By establishing a correlation between the external and internal datasets, this study reconstructed the real-time motion boundaries of the internal target area, achieving real-time monitoring of tumor volume (GTV) motion variations. The model's accuracy and computational efficiency were improved by narrowing the input body surface area. The model performance was evaluated using Euclidean distance, DICE coefficient, 95% Hausdorff distance (HD95), and mean surface distance (MSD).

Results:

In the model accuracy evaluation metrics, the mean DICE coefficient was 0.89±0.09, with a maximum value of 1 and a minimum value of 0.67. The mean MSD was 1.04±0.73 mm, with a minimum of 0.32 mm and a maximum of 1.92 mm. The mean HD95 was 2.69±2.32 mm, with a maximum of 5.28 mm and a minimum of 0.36 mm. The mean Euclidean distance between the predicted and actual centroid points was 1.48±1.14 mm, with a maximum of 2.66 mm and a minimum of 0.46 mm. Both external validation datasets had a DICE value greater than 0.85, specifically 0.88±0.12 and 0.86±0.12. The model’s accuracy when using a maximum deformable surface region with a 100 mm input radius showd not significantly different from that with the entire surface input. This indicates that most non-critical regions (i.e., body surface data beyond the 100 mm range) have negligible impact on the model's performance on the model's performance and can be ignored.

Conclusion:

The prediction model developed in this study, based on the 3D nnUnet deep learning framework, is capable of accurately predicting the real-time position of lung tumors (GTV) based on surface motion signals, exhibiting high levels of accuracy and robustness.