Keywords: Image Registration, Neural Networks, Quality Assurance
TL;DR: This project outlines a method of producing quantitative measures of registration performance in the absence of labels on one of the images.
Abstract: Patient-specific quality assurance of image registrations is needed to enable their use in adaptive radiotherapy. An automated method of assessing the quality of a registration between head and neck CT scans was investigated. Ground truth organ contours were propagated to the deformed image from the floating image and then compared with contours on the reference image by calculating Dice similarity. A 2D convolutional neural network was designed to predict Dice coefficients based on the reference image and the displacement vector field of the registration. The network was trained using axial slices of head and neck CT images to predict the metric for the spinal canal, parotid glands and brainstem. The network was able to predict the Dice coefficient for unseen images. For the spinal canal, 95%of predictions were within 0.208 of the true value, with an average absolute difference of 0.0811. For the left parotid gland, 95% were within 0.270 of the true value, with an average absolute difference of 0.0987. This demonstrates that convolutional neural networks can be trained to effectively predict similarity metrics which can be used to assess the quality of an automatically produced registration.
Code Of Conduct: I have read and accept the code of conduct.