A Hybrid Deep Learning and Reinforcement Learning Framework for Intelligent Error Detection in Radiotherapy Planning Systems

Abstract

Error prevention in radiotherapy planning is essential for effective tumour control with minimal toxicity to normal tissues. Even though clinical protocols are adhered to strictly, human and semi-automatic processes are susceptible to minute but crucial planning errors. This work introduces a novel hybrid strategy combining deep learning with reinforcement learning for error detection in planning. The deep learning module uses convolutional and attention-based architectures to learn dosimetric and spatial patterns from delineated anatomy, dose maps, and CT scans. The reinforcement learning agent learns to iteratively update error classification strategies with feedback from previous treatment planning data and expert corrections. The framework was validated using a carefully selected set of 1,200 radiotherapy plan annotations with errors categorized across dose deviation, target under-coverage, and organ-at-risk violation errors. Results in experiments showed the model achieving up to a maximum F1-score of 96.4% for binary classification and accuracy at 88.9% for multi-class categorization of error, together with a reduction in false negatives of 21.3% over standard deep neural networks. Additionally, policy learning further improved detection in complex multi-class cases. The performance gain itself translates directly into decreased risk of toxicity and better treatment accuracy for clinical applications. These findings indicate that an end-to-end deep reinforcement learning approach can provide a scalable and intelligent solution for radiotherapy quality assurance with considerable treatment safety and planning reliability improvements.

Ähnliche Arbeiten