Author: Lu Jiang, Ke Sheng 👨🔬
Affiliation: Department of Radiation Oncology, University of California at San Francisco, Department of Radiation Oncology, University of California, San Francisco 🌍
Purpose:
Conventional radiotherapy treatment planning is guided by a set of generic objectives that are unspecific to patient anatomy. Treatment planning thus heavily relies on the planner’s experience and available time, which often lead to inconsistent plan quality. Recently emerged deep learning methods show the promise of predicting the dose distribution based on individual anatomies, but leave space for improvement in prediction accuracy and interpretability. Here, we introduce the UKAN model, integrating Kolmogorov-Arnold Network(KAN) layers into the U-Net architecture to improve dose prediction accuracy and explainability.
Methods:
We adopt the interpretable UKAN architecture to predict the dose distribution for head and neck patients. The tokenized KAN modules replaced the last convolution layers of the U-Net, and the decoder retained the U-shaped architecture to ensure effective feature propagation for 3D voxel-level predictions. Experiments were conducted using the OpenKBP challenge dataset, which included 340 patients, and results were benchmarked against the standard U-Net model. Model performance evaluations included mean absolute error(MAE), structure similarity index(SSIM), peak signal-to-noise ratio(PSNR), as well as dose-specific metrics such as distance-to-agreement(DTA) and global gamma pass rate(%GP). Interpretability was analyzed and compared using post-hoc activated patterns.
Results:
UKAN demonstrated superior performance over basic U-Net, reducing MAE (2.490Gy vs. 2.590Gy), improving SSIM(0.846 vs. 0.842) and PSNR(26.166 vs. 25.973). DTA decreased by 8.64%(2.279mm vs. 2.494mm), and global %GP increased from 68.92% to 70.17%. UKAN showed a marked improvement in the ability to precisely locate the planning target volume(PTV) and activate the boundaries closely aligned with the ground truth PTV masks.
Conclusion:
By integrating KAN layers into the U-Net, the UKAN architecture improves the accuracy of 3D dose prediction for head and neck radiotherapy plans. The distribution of improvement shows the model awareness of the PTV boundary, which is the most important geometrical feature in treatment planning.