Dynamic stochastic deep learning approaches for predicting geometric changes in head and neck cancer

Area of Science:

• Radiotherapy
• Medical Physics
• Machine Learning

Background:

• Modern radiotherapy requires efficient adaptation to patient anatomical changes during treatment.
• Predicting these variations, such as organ deformation or tumor shrinkage, is essential for optimizing treatment plans.
• A framework to predict future patient states based on initial attributes and learned patterns is highly desirable.

Purpose of the Study:

• To investigate the feasibility of predicting patient anatomical changes (volume and setup variations) during fractionated radiotherapy.
• To compare the performance of a novel quantum mechanics-based recurrent neural network (QRNN) with a classical Markov-based recurrent neural network (MRNN).

Main Methods:

• Developed and evaluated two predictive frameworks: QRNN and MRNN.
• Utilized a dataset of 125 head and neck cancer patients, augmented with synthetic data from a generative adversarial network.
• Assessed model performance using area under the receiver operating characteristic curve (AUC) scores.

Main Results:

• The MRNN framework consistently outperformed the QRNN framework across various state vector sizes.
• MRNN achieved higher validation AUC scores (e.g., 0.742±0.021 for size 4) compared to QRNN (e.g., 0.675±0.036 for size 4).
• Statistically significant performance differences (p<0.05) were observed for higher-order state vectors.

Conclusions:

• Both MRNN and QRNN models show potential for predicting patient anatomical changes in adaptive radiotherapy.
• The MRNN approach, based on classical Markov processes, demonstrated superior predictive accuracy in this study.
• These findings support the use of stochastic models for enhancing adaptive radiotherapy planning.