Ravasio, Claudio S.;
(2022)
Deep Retinal Optical Flow: From Synthetic Dataset Generation to Framework Creation and Evaluation.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Ravasio_Thesis_Final.pdf - Published Version Download (22MB) | Preview |
Abstract
Sustained delivery of regenerative retinal therapies by robotic systems requires intra-operative tracking of the retinal fundus. This thesis presents a supervised convolutional neural network to densely predict optical flow of the retinal fundus, using semantic segmentation as an auxiliary task. Retinal flow information missing due to occlusion by surgical tools or other effects is implicitly inpainted, allowing for the robust tracking of surgical targets. As manual annotation of optical flow is infeasible, a flexible algorithm for the generation of large synthetic training datasets on the basis of given intra-operative retinal images and tool templates is developed. The compositing of synthetic images is approached as a layer-wise operation implementing a number of transforms at every level which can be extended as required, mimicking the various phenomena visible in real data. Optical flow ground truth is calculated from motion transforms with the help of oflib, an open-source optical flow library available from the Python Package Index. It enables the user to manipulate, evaluate, and combine flow fields. The PyTorch version of oflib is fully differentiable and therefore suitable for use in deep learning methods requiring back-propagation. The optical flow estimation from the network trained on synthetic data is evaluated using three performance metrics obtained from tracking a grid and sparsely annotated ground truth points. The evaluation benchmark consists of a series of challenging real intra-operative clips obtained from an extensive internally acquired dataset encompassing representative surgical cases. The deep learning approach clearly outperforms variational baseline methods and is shown to generalise well to real data showing scenarios routinely observed during vitreoretinal procedures. This indicates complex synthetic training datasets can be used to specifically guide optical flow estimation, laying the foundation for a robust system which can assist with intra-operative tracking of moving surgical targets even when occluded.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Deep Retinal Optical Flow: From Synthetic Dataset Generation to Framework Creation and Evaluation |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2022. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| Keywords: | Optical flow, computer vision, machine learning, synthetic data, oflib |
| UCL classification: | UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Med Phys and Biomedical Eng UCL > Provost and Vice Provost Offices > UCL BEAMS UCL |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10157675 |
Loading...
Loading...Loading...Loading...Archive Staff Only
 |
View Item |
