Free Papers

Automatic detection of any abnormal fundus photograph in the general population

Free Paper Details

First Author: B.Cochener-Lamard FRANCE

Co Author(s):    M. Lamard   J. Rottier   A. Le Guilcher   P. Massin   G. Quellec   S. Matta     

Abstract Details

Purpose:

The objective of this study is to adapt our previously developed artificial intelligence (AI) to an external dataset coming from a more general population, as part of mass screening in private practice. Our AI was originally trained on a Diabetic Retinopathy screening network (OPHDIAT, France) to automatically differentiate normal photographs from photographs showing at least one pathological sign. In this work, we generalize our AI towards a wider and different population, not only targeting diabetic retinopathy patients.

Setting:

1-University Brest Hospital - 5 Avenue FOCH 29609 Brest - France 2- Groupe Santé Sud CMCM, 28 Rue de Guetteloup,Le Mans, F-72100 France 3- Inserm LaTIM, UMR 1101, Brest, F-29200 France 4-Service d’Ophtalmologie, Hôpital Lariboisière, APHP, Paris, F-75475 France 5-Groupe Evolucare Technologies 60, route de Sartrouville, 78230 Le pecq

Methods:

During the reported period, 2017-2019, 8669 exams were conducted in the private screening network. The mean age of the population was 33 +- 22 years (minimum: 0, maximum: 97), and 56% were female. For each examination, only one ground truth annotation was obtained by a senior ophthalmologist. The report included the doctor's opinion: normal, abnormal or of poor quality. To adapt our AI to this dataset, 8131 examinations (17120 Images) of sufficient quality were selected. Approximately 50 % of them (8583 Images) were used for testing. The remaining were used for tuning (6831 Images) and validation (1706 Images).

Results:

Without adaptation to the new dataset, our AI is able to differentiate between normal and abnormal examinations with an area under the curve (AUC) of 0.8347 on the test set. It can detect 60% of pathological cases with a specificity 90%. After tuning our AI using the private network dataset, retinal pathologies were correctly identified with an area under the ROC curve of 0.9086. The algorithm can detect 79% of pathological cases with a specificity 90%.

Conclusions:

Automatic mass eye pathology screening in the general population is vital for preventing visual impairment. The adaptation of our AI, initially trained on a diabetic population, to screen retinal pathologies on a private network is very promising. The results show that the automatic screening performance is greatly improved when tuning the algorithm with the private network dataset which has different characteristic (general population) and image interpretation criteria. The algorithm has higher sensitivity and AUC score enabling it to generalize better. the way is open for automatic sorting between normal and abnormal retina on any type of database.

Financial Disclosure:

... research is funded, fully or partially, by a competing company