Intra-operative OCT (iOCT) Image Quality Enhancement: A Super-Resolution Approach using High Quality iOCT 3D Scans
Presented at Ophthalmic Medical Image Analysis (OMIA) 2021, MICCAI WORKSHOP: Intra-operative OCT (iOCT) Image Quality Enhancement
This reposirtory contains the implementation of the methods presented in the paper "Intra-operative OCT (iOCT) Image Quality Enhancement: A Super-Resolution Approach using High Quality iOCT 3D Scans", presented at OMIA 2021.
Effective treatment of degenerative retinal diseases will require robot-assisted intraretinal therapy delivery supported by excellent retinal layer visualisation capabilities. Intra-operative Optical Coherence Tomography (iOCT) is an imaging modality which provides real-time, cross-sectional retinal images partially allowing visualisation of the layers where the sight restoring treatments should be delivered. Unfortunately, iOCT systems sacrifice image quality for high frame rates, making the identification of pertinent layers challenging. This paper proposes a Super-Resolution pipeline to enhance the quality of iOCT images leveraging information from iOCT 3D cube scans.We first explore whether 3D iOCT cube scans can indeed be used as high-resolution (HR) images by performing Image Quality Assessment. Then, we apply non-rigid image registration to generate partially aligned pairs, and we carry out data augmentation to increase the available training data. Finally, we use CycleGAN to transfer the quality between LR and HR domain. Quantitative analysis demonstrates that iOCT quality increases with statistical significance, but a qualitative study with expert clinicians is inconclusive with regards to their preferences.
We used an internal dataset of intra-operative retinal surgery videos and OCT/iOCT scans from 66 patients acquired at Moorfields Eye Hospital, London, UK. We ended up having 983 images per type (V-iOCT, C-iOCT, C-pOCT).
IQA metrics implementation files can be found in metrics
. For performing IQA, run the metrics/main.py
and specify the two folder paths that contain the images you want to compare calculating
L_feat, FID (CleanFID), GCF. NIQE score can be calculated by running niqe_score.m
Registration implementation files can be found in registration
. Registration is performed using SimpleElastix.
Installation is required to run the registration/main.py
.
For registration:
registration/main.py -t registration -fp path_to_fixed -mp path_to_moving -sp path_to_moved
For validation of the registration:
registration/main.py -t validation -fp path_to_fixed -sp path_to_moved
For data augmentation:
registration/main.py -t augmentation -sp path_to_moved -v videofile
For registration of augmented data:
registration/main.py -t reg_augmentation -fp path_to_fixed -mp path_to_moving -sp path_to_moved
For CycleGAN model we used the [Tensorflow] implementation by Xiaowei Hu.
For Pix2Pix we used the [Tensorflow] implementation by Christopher Hesse.
Table 1 summarises IQA results. The values of the three metrics (FID, l_feat, NIQE) are lower for C-iOCT which indicates better perceptual quality with respect to C-pOCT and thus they can be used as HR images.
The analysis was based on the: V-iOCT (LR), SR-iOCT(CGAN) (SR using CycleGAN), SR-iOCT(Pix) (SR using Pix2Pix) and C-iOCT (HR) test sets. FID, l_feat, NIQE and the no-reference Global Contrast Factor (GCF) metrics were used.
If you use this code for your research, please cite our paper:
@inproceedings{komninos2021intra,
title={Intra-operative OCT (iOCT) Image Quality Enhancement: A Super-Resolution Approach Using High Quality iOCT 3D Scans},
author={Komninos, Charalampos and Pissas, Theodoros and Flores, Blanca and Bloch, Edward and Vercauteren, Tom and Ourselin, S{\'e}bastien and Da Cruz, Lyndon and Bergeles, Christos},
booktitle={8th International Workshop on Ophthalmic Medical Image Analysis, OMIA 2021 held in conjunction with 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021},
pages={21--31},
year={2021},
organization={Springer Science and Business Media Deutschland GmbH}
}