Table 1 Technical and image-related challenges to development of deep learning algorithms for ocular disease detection.
Challenges | Research question | Paper addressing this question | Answer to research question | |
|---|---|---|---|---|
Technical | Newer convolutional neural networks with increasing number and complexity of layers may allow for greater depth of analysis but may intensify burden on hardware processing power and memory. | Does altering the convolutional neural network architecture affect performance? | Current paper | No. Different neural networks do not affect performance. |
Differences between computational frameworks based on flexibility, applicability, speed, ease of use, may affect choice. | Does altering the computational framework affect performance? | Current paper | No. Different computational frameworks do not affect performance. | |
Image-related | Lack of access to high quality retinal images due to poor fundus camera specifications, reduced storage space, or compression for tele-ophthalmology. | Does altering the level of compression of the input data affect performance? | Current paper | Yes. Reducing image size below 250 KB drops performance significantly. |
Different groups in various countries may possess datasets with varying number of field of fundus views due to disparities in protocols, resources, and manpower. | Does altering the number of fundus field of views of the input data affect performance? | Current paper | Yes. Performance drops in descending order from 7-field to 2-field to 1-field. | |
The presence of cataract may impinge on proper visualization of the fundus and inaccurate diagnosis due to media opacity, light scatter and aberrancies. | Does previous cataract surgery affect performance? | Current paper | Yes. Presence of media opacity in phakic eyes reduces performance. | |
The range of retinal cameras available to capture fundus images in terms of camera specifications, requirement for mydriasis, may provide variability in degree of field of view and image quality output. | Does altering the retinal cameras used affect performance? | Ting et al.15 | No. Different retinal cameras do not affect performance. | |
Ethnic differences in eyes exist that affect optical systems’ ability to capture the posterior pole and the identification of the norm (e.g. pigmentation, optic disc size, vasculature). | Do images from various ethnic groups affect performance? | No. Images from different ethnic groups do not affect performance. | ||
Different populations vary in prevalence rates of ocular disease, thus affecting the dataset used for validation and the utility of a clinical test deployed in that population. | Does deployment in populations with different disease prevalence rates affect performance? | No. Deployment in populations with different prevalence rates does not affect performance. | ||
Ocular diseases do not develop distinctly as many share similar risk factors and occur concurrently in the same patient, thus distinction between manifestations of different diseases is paramount. | Does concurrent related ocular diseases affect performance in detection of an individual disease? | Ting et al.15 | No. Other existing diseases do not affect the algorithm’s ability to detect individual diseases accurately. | |
The type of study (population-based, clinic-based or screening cohort) used to collect retinal images may influence the patient demographics of the datasets. | Does the type of studies affect the performance? | Ting et al.15 | No. The type of study does not affect the performance. | |
Different countries may use different reference standards for grading of diabetic retinopathy (e.g., grader or ophthalmologist), a product of resource allocation, expertise and training available. | Does the difference in reference standard used for labeling of images affect performance? | Ting et al.15 | No. Different reference standards used do not affect the performance. | |
Availability of large datasets in the target population may be scarce and insufficient for the training required for a highly performing algorithm. | Does a smaller dataset used for training affect the performance? | Yes. Datasets that drop below 60,000 images produce large drops in performance. | ||
With large amount of images required for training, time constraints and reduced access to high quality retinal cameras may limit the use of large high resolution images for training of deep learning systems. | Does image size of the training dataset affect the performance? | Sahlsten et al.25 | Yes. Increased resolution of training images produce better performance but increases training time. | |
Mydriasis may provide greater visualization for photographic capture of the posterior pole, potentially influencing quality of fundus photographs. | Does mydriatic photographs improve performance compared to non-mydriatic images? | No. Mydriasis does not significantly improve performance. | ||
