Ade Akintonde & Shaheer Saeed– CMIC/WEISS joint seminar series

Speaker: Ade Akintonde

Title: Surrogate driven respiratory motion model derived from CBCT projection data

Abstract: Cone beam CT is the most common imaging method for image guided radiotherapy. However due to the slow rotating gantry, the image quality of CBCT can be adversely affected by respiratory motion, as it blurs the tumour and nearby organs at risk, which makes visualization of organ boundaries difficult, in particular for organs in the thoracic region. Currently one approach to tackle the problem of respiratory motion is the use of respiratory motion model to compensate for the motion during CBCT image reconstruction. The overall goal of this work is to estimate the 3D motion, including the breath-to-breath variability, on the day of treatment directly from the CBCT projection data, without requiring any external devices. The work presented here consist of two main parts: firstly, we introduce a novel data driven method based on principal component analysis with the goal to extract a surrogate signal related to the internal anatomy, from the CBCT projections. Secondly, using the extracted signals, we use surrogate-driven respiratory motion models to estimate the patient’s 3D respiratory motion. We utilized a recently developed generalized framework that unifies image registration and correspondence model fitting into a single optimization. This enables the model to be fitted directly to unsorted/unreconstructed data (CBCT projection data), thereby allowing an estimate of the patient’s respiratory motion on the day of treatment. To evaluate our methods, we have used an anthropomorphic software phantom combined with CBCT projection simulations. We have also tested the proposed method on clinical data with promising results obtained.

Speaker: Shaheer Saeed

Title: Learning image quality assessment by reinforcing task amenable data selection

Abstract: In this talk,we consider a type of image quality assessment as a task-specific measurement, which can be used to select images that are more amenable to a given target task, such as image classification or segmentation. We propose to train simultaneously two neural networks for image selection and a target task using reinforcement learning. A controller network learns an image selection policy by maximising an accumulated reward based on the target task performance on the controller-selected validation set, whilst the target task predictor is optimised using the training set. The trained controller is therefore able to reject those images that lead to poor accuracy in the target task. In this work, we show that the controller-predicted image quality can be significantly different from the task-specific image quality labels that are manually defined by humans. Furthermore, we demonstrate that it is possible to learn effective image quality assessment without using a ``clean'' validation set, thereby avoiding the requirement for human labelling of images with respect to their amenability for the task. Using 6712, labelled and segmented, clinical ultrasound images from 259 patients, experimental results on holdout data show that the proposed image quality assessment achieved a mean classification accuracy of 0.94±0.01 and a mean segmentation Dice of 0.89±0.02, by discarding 5% and 15% of the acquired images, respectively. The significantly improved performance was observed for both tested tasks, compared with the respective 0.90±0.01 and 0.82±0.02 from networks without considering task amenability. This enables image quality feedback during real-time ultrasound acquisition among many other medical imaging applications.

Elevator pitch speakers: Bojidar Rangelov & Zhe Min

Chair: Jamie McClelland

Link to the Moodle page here: https://moodle.ucl.ac.uk/course/view.php?id=19613   

Please enrol with key: CMIC