AIROGS Lite¶
Motivation¶
- Early detection of glaucoma can avoid visual impairment, which could be facilitated through screening.
- Artificial intelligence (AI) could increase the cost-effectiveness of glaucoma screening, by reducing the need for manual labor. AI approaches for glaucoma detection from color fundus photography (CFP).
Aim¶
- Development an AI solution for glaucoma screening by classifying CFPs as “referable glaucoma” or “no referable glaucoma”.
Data¶
- Development set* (15,000 CFPs)
- 13,500 "No referable glaucoma" (NRG) CFPs
- 1,500 "Referable glaucoma" (RG) CFPs
- Test labels will not be available, only the images. You will need to upload your predictions to our challenge website.
*The use of additional CFP development data (including other data from the original AIROGS challenge and weights pretrained on fundus image data) is prohibited. The use of such data or pretraining is also not allowed in any stage of the algorithm (so also not in a preprocessing or postprocessing step). The use of other data and pretraining with other data, such as natural images such as those from ImageNet or other medical images is allowed.
Test set evaluation¶
- Excepted output: Glaucoma Likelihood (as CSV file).
- Metrics: Partial AUROC (90-100% specificity) and Sensitivity @ 95% specificity. A ranking will be made for these two metrics. The mean of those rankings is the final score. Implementation of the two metrics can be found here and here.
- The RG prevalence in the test set will be similar to the train set.
- The test set images was released on October 10, 2022. You can download test_cfp.zip (the test images) and submission_example.csv (an example submission file) here: https://zenodo.org/record/7178671.
- Three submissions to the test set will be allowed. You will see the results (almost) immediately.
The approach of choice is up to you. Method justification, experimentation and reporting are more important than a high score!