Criteria-calibration approaches to deep learning-based cervical cancer radiation treatment auto-planning.

BACKGROUND

Knowledge-Based Planning (KBP) pipelines, which integrate machine learning-based models to predict dose distribution, have gained popularity in clinical radiation therapy. However, for patients with specific requirements, the trained models may struggle to rapidly adjust to guide the automatic planning process. Therefore, the aim of this study was to calibrate the dose prediction model to improve the quality and accuracy of automatic planning for cervical cancer radiation therapy.

MATERIALS AND METHODS

We retrospectively collected a routine cervical cancer dataset (200 cases) to conduct the KBP pipelines for automatically generating radiation planning, and a small number of ovarian-protection and myelosuppressive datasets (21 cases) to calibrate and evaluate the dose prediction model. A total of three criteria-calibration approaches to solve the data imbalance problem in dose prediction were introduced and compared, including Prediction Tolerance function on uTPS (United Imaging Healthcare Co., Ltd., Shanghai), transfer learning, and mixture density network.

RESULTS

The Prediction Tolerance function allowed for rapid optimization adjustments without model modification, which is suitable for patients with strong desires for ovary protection. The transfer learning approach required minimal training time and data to generate acceptable automatic planning results. The Mixture Density Network (MDN) approach, although the most time-consuming to train, achieved robust prediction results and facilitated dataset analysis. The MDN method showed the greatest consistency between predicted dose distribution and actual optimization outcomes, highlighting its potential as a reliable calibration method for dose prediction.

CONCLUSION

This study demonstrated an automatic KBP workflow and compared three criteria-calibration approaches to address the data imbalance problem in dose prediction. These approaches can partially calibrate pre-existing models to accommodate newly added criteria and could be implemented according to specific requirements in different scenarios. Although there are trade-offs in various aspects, they all can generate feasible radiation treatment plans.