Magnetic Resonance (MR)-guided online adaptive radiotherapy has the potential to revolutionise cancer treatment. It exploits soft-tissue contrast of MR images obtained right before patient’s radiation treatment to personalise radiotherapy treatment plans. Four-dimensional (4D) MR Imaging (MRI) is used to adapt radiation to maximally target the tumour while sparing healthy tissues; accurate tumour location is essential so that high radiation dose can be applied to speed up patient recovery compared to conventional radiotherapy treatment.
This research builds on advances in adaptive radiotherapy [1] to explore neuro-symbolic approaches for MRI data processing. Convolutional Neural Networks (CNNs) are used in medical imaging [2] but have two drawbacks: no native support for uncertainty estimation, and no explanation of how predictions are obtained. Such drawbacks can lead to serious consequences for online adaptive radiotherapy, since high radiation dose is intended for tumour regions identified with low uncertainty.
This project concerns improving trustworthiness of neural-network based medical image processing by addressing these drawbacks. We aim to pioneer a new approach that combines two recent advances: Bayesian Neural Networks (BNNs) which support uncertainty estimation of predictions, and Emergent Symbolic Language (ESL) which shows promise in interpretable capabilities; both have been used in medical image processing [3], [4]. Our research will cover three themes: approach development, application adaptation, and effective optimisations.
The first theme focuses on a new approach for improving trustworthiness of neural networks by combining BNNs and ESL. Since ESL involves deep neural networks and symbol generator [4], one possibility is to replace the neural networks by BNNs to provide uncertainty estimates for the results. This study will include tuning the combined configuration to deliver uncertainty estimation for the generated symbols and the final decisions. We will start with extending 2D and 3D BNNs in our previous work [5] to cover ESL. Other promising neuro-symbolic approaches capable of dealing with uncertainty, such as Bayesian-symbolic physics [6], will be studied and compared with our approach.
The second theme adapts the work in the first theme to support 4D MRI data processing. We will collaborate with medical imaging experts who would confirm, for example, that the uncertainty estimates are as expected to demonstrate improvement of trustworthiness. Other potential benefits, such as prevention of over-fitting and reduction in the number of training samples, will be explored. Case studies involving deep learning for 4D MRI [1], [7], such as image registration [8] and image segmentation [9], will be selected to show how trustworthiness can be improved using public domain data.
The third theme investigates effective algorithmic and implementation optimisations to determine the best ones for the proposed approach, to support online data processing while preserving trustworthiness. Since BNNs involve additional computations such as marginalisation, their training and inference procedures can be slower than CNNs with similar network structures. This study explores optimisations such as quantisation and multi-exit strategies for enhancing performance. The objective is to develop novel methods and tools to automate the development of an optimised implementation with given performance and accuracy that minimises the uncertainty of given symbols.