Visual innovation through AI in ophthalmology
Optical coherence tomography (OCT) is a gentle laser microscope and is used by ophthalmologists to examine the retina for abnormalities. Trained AI models support doctors in recognizing pathologies.
-
Initial situation
OCT scans can be used to record the fine layered structure of the retina. However, evaluating 3D scans for tumors, for example, is time-consuming and has not been tested much to date. Artificial intelligence (AI) can support doctors in this process.
-
Solution SCS
SCS trained a modified U-Net-CNN model for the detection of eye layers and tumors. The accuracy achieved is equivalent to that of experienced ophthalmologists.
The model runs dockerized as a service on an on-premise platform. -
Added value
The AI solutions automate and accelerate the detection of eye layers and tumors in 3D OCT scans. The model works for various OCT devices. This means that the AI can be used by many ophthalmologists in research.
Detection of defects in the eye
Due to its intrinsic 3D resolution and non-invasive ease of use, OCT is widely used in clinics to diagnose major eye diseases, including age-related macular degeneration, glaucoma, diabetic retinopathy and vitreo-retinal traction.
Artificial intelligence (AI) can help doctors to detect these pathologies in OCT scans more efficiently. The AI-supported findings are evaluated by doctors and shorten the time a doctor spends in front of the computer. And they potentially reduce the diagnostic error rate.
Label data
Labeling the eye layers requires a great deal of expertise - and a lot of time. Several experts labeled hundreds to thousands of OCT-B scans pixel by pixel. SCS stored these in versioned form.
Decision in favor of the architecture of the modified U-Net
As a basis, research was carried out into the CNN architectures used for semantic segmentation of medical images. The U-Net has proven itself and still achieves state-of-the-art results, i.e. the resulting error corresponds to the irreducible error. In other words, the U-Net is indistinguishable in its accuracy from ophthalmology experts.
Results of segmentation without tumor images
The project comprises several studies and evaluations. The one described below is limited to an illustrative evaluation without tumor images. The "human factor" plays a major role in the results. Even humans do not produce a 100% (see line 1 in the table below) consistent ground truth (GT). It therefore makes sense to compare the variations in segmentation between different people. Hamming Distance was used as a measure for comparing performance. The U-Net performs better than the heterogeneous group of labelers consisting of laypersons, opticians and professional ophthalmologists.
| Comparison | Variation |
|---|---|
| Within GT (1 expert, 3 times repeated labeling) | 1.3-1.7% |
| GT and CNN | 1.4-1.7% |
| 10 people (amateurs, opticians, ophthalmologists) | 1.7-3.5% |
Reproducibility
The software code is managed in a Git repository. Data sets are versioned and stored "immutably" on a network drive. Experiments are kept in a digital lab journal and final versions (for publication[i] or production) are also versioned. Reproducibility is extremely important in this project, as the results are used for publications.
Rollout and operation
The backend runs in an on-premise infrastructure in dockerized containers and will be available to the medical community in the future. Users upload queries including their anonymized OCT scans via a user-friendly web interface. The model then determines the volume and areas of the eye layers and makes the results available to the user via a web interface.
Few-Shot Learning
A particular challenge is the small amount of tumor data within the images. In another study, various techniques for the artificial generation of ground truth data were tested. Approaches using autoencoders delivered blurred results. Only the use of Generative Adversarial Networks (GANs) allowed a reliable artificial generation of tumor images.
Summary
Provided that sufficient, i.e. hundreds to thousands of high-quality and pixel-by-pixel labeled OCT scan images are available, the U-Net recognizes medically relevant structures with an accuracy comparable to human ophthalmology experts.
[i] Validation of automated artificial intelligence segmentation of optical coherence tomography images
Maloca PM, Aaron Y Lee, Emanuel R de Carvalho, Mali Okada, Katrin Fasler, Irene Leung, Beat Hörmann, Pascal Kaiser, Susanne Suter, Pascal W Hasler, Javier Zarranz-Ventura, Catherine Egan, Tjebo FC Heeren, Konstantinos Balaskas, Adnan Tufail, Hendrik PN Scholl Scholl (2019) Validation of automated artificial intelligence segmentation of optical coherence tomography images. PLOS ONE 14(8): e0220063