Author: Yan Dai, Jie Deng, Christopher Kabat, Weiguo Lu, Ying Zhang, Hengrui Zhao 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose:
MRI-guided adaptive radiotherapy (MRgART) using MR-LINAC systems offers significant advantages for liver cancer, enabling superior tumor delineation and online plan adaptation. However, manual liver tumor contouring is time-intensive (15-20 minutes), delaying the online ART workflow. Longitudinal MRI data from simulation and prior treatments contain patient-specific information that can expedite contouring. This study proposes a novel approach integrating patient-specific prior knowledge to enhance liver tumor auto-contouring.
Methods:
Data from 48 liver cancer patients (2-6 fractional MRIs per patient) treated with MRgART on a 1.5T MR-LINAC were collected, including 207 target volumes (GTV, ITV, CTV) and planning images. Patients were split into training (34), validation (9), and test (5) sets. A Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) units embedded in U-Net skip connections was developed to capture multi-level memory representations of historical knowledge. The network used 3 input channels (planning image, registered prior image, and registered prior target) and 4 down-sampling layers. All fractional MRIs for each patient were sequentially processed during training and testing.
Results:
The RNN model achieved a Dice score of 69.2%, outperforming co-registration (68.1%) and conventional U-Net (67.1%). In optimal cases, the Dice score improved by up to 4.2%, attributed to the LSTM’s ability to retain historical information beyond the latest prior knowledge.
Conclusion:
The proposed model demonstrates potential to accelerate target contouring, streamlining MR-LINAC adaptive planning. Future work leveraging pancreatic and brain datasets to improve auto-segmentation for drastically changing targets is warranted.