Project Details
Description
Project Summary
Retinopathy of prematurity (ROP) is a leading yet largely preventable cause of childhood blindness globally.
There is an unmet need for timely and consistent diagnosis of treatment-required ROP, such as plus disease of
ROP. A deep learning-based diagnostic tool would be advantageous to reduce the dependence on clinical
expertise during the diagnostic phase. However, there are two significant challenges in training deep learning-
based approaches. The first challenge is that the performance of deep learning-based diagnosis is hindered by
limited data availability in the training phase. The size of the training data set directly influences training efficiency
and the ability to generalize. Since premature babies are vulnerable populations, images from these babies are
inherently limited and insufficient to cover all potential scenarios of ROP for training. The second challenge is
the variation in image quality due to differences in acquisition systems. The differences in quality data sets will
impact the deep learning model. They will likely fail to produce accurate diagnoses due to the lack of alignment
between the training data and data acquired from other platforms. Therefore, there is a great need for new
methods and technology to increase the training set and improve or normalize the image quality to develop a
diagnostic deep learning model for ROP. To address this critical unmet clinical need, we have developed
technology to augment the training dataset using animal-to-human (A2H) translated images and implement
quality control by cross-platform retinal image standardization process (CRISP). In this study, we will
demonstrate that our A2H translated images will exhibit retinal vascular patterns of plus disease of ROP. These
translated images can be used to train the deep learning-based ROP diagnostic model. We will demonstrate that
our CRISP model can resolve the difference between different imaging acquisition systems and/or overall image
quality. We will evaluate the impact of quality control on the performance of ROP diagnosis. Specific Aim 1 is to
optimize and test an animal-to-human model using an existing retinal image dataset. Specific Aim 2 is to acquire
a robust set of longitudinal animal ROP images using two animal models and three different imaging platforms
and manually annotate the images for the training. Specific Aim 3 is to develop and test a CRISP model on
quality control. Specific Aim 4 is to validate the A2H model to diagnose ROP and evaluate the combination of
A2H and CRISP as a universal tool to improve learning performance. Our proposed method could potentially
improve the diagnosis of ROP. Furthermore, our innovative approach of using animal data to enhance human
studies could also be applied to other clinical studies where human data and clinical labeling resources are
limited. Finally, our quality control tool that can standardize the image quality will allow opportunities for more
robust image analysis and potentially applicable in telemedicine.
Status | Active |
---|---|
Effective start/end date | 1/09/24 → 31/08/25 |
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.