Fig 1.
Flowchart of our AI design, implementation and test.
The public data attaiment and upload onto the Microsoft Azure cloud platform was the first step. “quality assessment” CNN was trained to identify adequate and inadequate images. the entire public dataset was then devided to training, validation and test sets. The test set was then ‘curated’ by the “quality assessment” CNN. The DRCNN was trained on un-curated data, and then tested on ‘curated’ and ‘un-curated’ data. Its performance was also assessed using 2 or 3 DR labels.
Fig 2.
Samples of ‘adequate’ and ‘in-adequate’ images as decided by a senior retinal specialist.
Fundus images deemed adequate are shown in the upper row. Fundus images deemed inadequate are shown in the bottom row.
Table 1.
Re-categorization of the original Kaggle EyePACS grading scheme (5 grades) to three new categories.
Table 2.
Re-categorization of the original Kaggle EyePACS grading scheme (5 grades) to two new categories.
Fig 3.
Contrast enhancement of the Kaggle EyePACS fundus image.
The Gaussian blur technique was applied to the raw fundus image (left). This technique minimizes intensity and contrast variability in fundus image dataset (right).
Fig 4.
Training and validation process of DRCNN, with training cross-entropy loss (A), training accuracy (B), validation cross-entropy loss (C) and validation accuracy (D) are presented.
Fig 5.
Screenshot of the Microsoft Azure™ virtual machine.
A Virtual Machine was created on Microsoft Azure East US server. 6 CPUIs were avaialble to us on this Virtual Machine, and it was used for training and validation process.
Table 3.
Performance of the DRCNN based on three grades (healthy, non-referable, referable).
Table 4.
Performance of the DRCNN based on two grades (healthy, diseased).
Table 5.
Sensitivity boost of the ‘curated’ dataset with three labels, for healthy and diseased categories.
Table 6.
Sensitivity boost of the ‘un-curated’ dataset with three labels, for healthy and diseased categories.
Table 7.
Sensitivity boost of the ‘curated’ dataset with two labels, for healthy and diseased categories.
Table 8.
Sensitivity boost of the ‘un-curated’ dataset with two labels, for healthy and diseased categories.
Table 9.
Comparison of un-boosted and boosted DRCNN with previous studies, which used the EyePACS dataset.
The highlighted cells show the ‘performance gains’ due to different boosting strategy.