Developments In Deeptuning: Translational Invariance of Trade-Off Feature Extraction In Real-Time Tunable Dose Prediction Framework
Abstract
Purpose
DeepTuning is a deep learning–based dose prediction framework that generates multiple dose distributions with different trade-offs, analogous to multi-criteria optimization. Leveraging historical cases with varying trade-offs, DeepTuning extracts semantic trade-off features from prior dose distributions and anatomical contours. These features can then be injected into the prediction pipeline for new patients, enabling physicians to review multiple dose distributions and explore trade-offs in real time. Because trade-offs between OARs are invariant to global translations of an image, the neural representation of the trade-off should be invariant as well. This study investigates the translational invariance of trade-off features extracted by DeepTuning.
Methods
DeepTuning adopts a U-Net architecture with a dual encoder and a shared decoder design. A posterior encoder extracts trade-off features from historical dose distributions, while a prior encoder captures patient-specific anatomical geometry from contours. Trade-off features are injected at the U-Net bottleneck, and the shared decoder generates dose distributions with corresponding trade-offs. Two model variants were implemented using different down-sampling strategies: (1) an anti-aliasing down-sampling approach consisting of a low-pass filter followed by strided-convolution, and (2) a conventional down-sampling approach using strided-convolution alone.
Results
Under spatial shifts of input images, the conventional model exhibited substantial variations in extracted trade-off features due to aliasing effects introduced during down-sampling. The variation of trade-off features reached up to 0.007 (measured by mean-square-error), resulting in distorted dose predictions. In contrast, the anti-aliasing model reduced trade-off feature variation to 0.001 and produced consistent dose predictions across spatial translations.
Conclusion
DeepTuning enables generation of real-time tunable dose distributions, allowing physicians to explore trade-offs upfront, reducing iterative communication with dosimetrists. Ensuring translational invariance of extracted trade-off features is critical for robust generalization across patient positioning. Incorporating anti-aliasing down-sampling effectively mitigates aliasing effects and improves the translational invariance of trade-off feature extraction and dose prediction.