Deep Learning Dose Optimization with Uncertainty Quantification for Intensity-Modulated Radiotherapy: A 3D Radiomics Approach

Abstract

Intensity-modulated radiotherapy planning demands complex optimization, balancing tumor control against normal tissue toxicity. This research introduces a hybrid deep learning framework combining 3D convolutional neural networks with radiomics features for automated dose distribution prediction. The architecture integrates Monte Carlo dropout and heteroscedastic regression to provide comprehensive uncertainty quantification, addressing critical gaps in clinical decision support systems. Evaluation of 340 head and neck cancer patients demonstrates mean absolute errors below 2.8% for planning target volumes and 3.1% for organs at risk, with gamma analysis pass rates exceeding 95.2% at 2 mm/2 % criteria. A comparative analysis across U-Net, ResNet, and DenseNet architectures establishes the superiority of radiomics-enhanced approaches, achieving 12.3% improvement in the dose conformity index and 18.7% reduction in prediction uncertainty compared with baseline methods. The uncertainty quantification provides clinically actionable confidence intervals supporting case triage and quality assurance prioritization while maintaining computational efficiency compatible with clinical workflows.