1409P - Advancing esophageal cancer radiotherapy: AS-NeSt's 3D predictive proficiency for personalized dose distribution

Background

Esophageal cancer treatment planning is challenged by the tumor's intricate nature and varying radiation prescriptions, compounded by the scarcity of robust predictive models. This study introduces the Asymmetric ResNeSt (AS-NeSt) model, a pioneering deep learning framework designed to predict 3D radiotherapy dose distributions across diverse prescription, significantly enhancing treatment precision and efficiency.

Methods

Using data from 530 esophageal cancer patients, we developed the AS-NeSt model, marking a pioneering advancement in radiation dose prediction. The model's core is its asymmetric encoder-decoder architecture, integrated with 3D ResNeSt blocks, ensuring high fidelity in detail preservation and data fitting. A novel MSE-based loss function enables precise optimization of dose distributions for diverse treatment plans, including single-target and simultaneous integrated boost (SIB) prescriptions.

Results

The AS-NeSt model demonstrated superior accuracy, maintaining prediction errors below 5% across all metrics, with a notable dice similarity coefficient of 0.93 for isodose volumes. It outperformed existing models, including HD-Unet, DoseNet, and DCNN, in accuracy, parameter efficiency, and prediction speed. Clinically, it facilitated a more accurate pre-treatment assessment, reduced planning time by over 50%, and minimized dosimetrist discrepancies, marking a significant step forward in personalized cancer treatment planning.

Conclusions

AS-NeSt represents a breakthrough in the application of deep learning for radiotherapy planning in esophageal cancer. By accurately predicting 3D dose distributions for various prescription types, it promises to streamline treatment planning, enhance outcome predictability, and optimize personalized care strategies. This model sets a new benchmark in integrating Artificial Intelligence (AI) into oncological treatment planning, offering a scalable solution for complex cancer types beyond esophageal carcinoma.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Y. Duan.

Funding

National Natural Science Foundation of China (12375346).

Disclosure

All authors have declared no conflicts of interest.