Page 266 - Computational Retinal Image Analysis

264 CHAPTER 13 Drusen and macular degeneration

Kermany et al. adapted an Inception V3 architecture pre-trained on the ImageNet
dataset in order to take advantage of the benefits offered by transfer learning [97].
108,312 images out of 207,130 OCT images (37,206 with choroidal neovasculariza-
tion, 11,349 with diabetic macular edema, 8617 with drusen, and 51,140 normal)
from 4686 patients passed an initial image quality review and were used to train the
AI system. The model was tested with 1000 images (250 from each category) from
633 patients. An accuracy of 96.6% with a sensitivity of 97.8% and specificity of
97.4% were achieved for the classification of choroidal neovascularization, diabetic
macular edema, drusen, and normal images. In both studies, an occlusion test was
used to highlight regions in images that contribute to the classification, which makes
it easier to understand how the network has “seen” the images and made the decision.
The recent work by DeepMind has adopted a different strategy [98]. First of
all, features such as layers and fluid spaces in OCT images are segmented based
on U-Net [99], then a CNN is used for the diagnosis of different conditions. They
used a population-based dataset from the Moorfields Hospital in London and this has
demonstrated the great potential of close collaboration between academia, NHS and
industry. However, the data is not publicly available.

5 Datasets

The (lack of) availability of datasets in medical research has been a considerable
challenge, but the recent open access policy requirements have made a difference in
this field. Below, we have listed some publicly available datasets for the research in
AMD and provided a brief description of them.

1. ARIA is a publicly available dataset in Liverpool, available at https://
eyecharity.weebly.com/aria_online.html. The collection includes 101 AMD
images and 60 non-AMD images, which were manually pre-labeled. All
images were taken using a fundus camera (FF450+, Zeiss Meditec, Inc.,
Dublin, CA) at a 50° field with a resolution of 576 × 768 pixels.
2. The STARE dataset (http://www.ces.clemson.edu/~ahoover/stare) comprises
97 images (59 AMD and 38 normal) taken using a fundus camera (TOPCON
TRV-50; Topcon Corp., Tokyo, Japan) at a 35° field and with a resolution of
605 × 700 pixels.
3. Kermany et al. [97] have made their OCT data available in JPEG format
through the public Mendeley database (https://doi.org/10.17632/rscbjbr9sj.3).
The images are split into those showing CNV, DME, drusen and normal
images. We initially obtained 207,130 OCT images. 108,312 images (37,206
with choroidal neovascularization, 11,349 with diabetic macular edema, 8617
with drusen, and 51,140 normal) from 4686 patients passed initial image
quality review and were used to train the AI system.
4. The Age-Related Eye Disease Study (AREDS) is a long-term, multi-center,
prospective study of 4757 people aged 55–80, supported by the National Eye
Institute and the National Institute of Health. The study involved cataract and