Detection of oedema on optical coherence tomography images using deep learning model trained on noisy clinical data. Issue 1 (15th May 2021)

• Record Type: Journal Article
• Authors: Potapenko, Ivan; Kristensen, Mads; Thiesson, Bo; Ilginis, Tomas; Lykke Sørensen, Torben; Nouri Hajari, Javad; Fuchs, Josefine; Hamann, Steffen; la Cour, Morten
• Is Part Of: Acta ophthalmologica. Volume 100:Issue 1(2022)
• Page Start: 103
• Page End: 110
• Publication Date: 2021-05-15
• Subjects: age related macular degeneration -- deep learning -- inter‐grader agreement -- retinal oedema -- training label noise; Ophthalmology -- Periodicals; 617.7005
• DOI: 10.1111/aos.14895
• Languages: English
• ISSNs: 1755-375X

Abstract

Abstract: Purpose: To meet the demands imposed by the continuing growth of the Age‐related macular degeneration (AMD) patient population, automation of follow‐ups by detecting retinal oedema using deep learning might be a viable approach. However, preparing and labelling data for training is time consuming. In this study, we investigate the feasibility of training a convolutional neural network (CNN) to accurately detect retinal oedema on optical coherence tomography (OCT) images of AMD patients with labels derived directly from clinical treatment decisions, without extensive preprocessing or relabelling. Methods: A total of 50 439 OCT images with associated treatment information were retrieved from databases at the Department of Ophthalmology, Rigshospitalet, Copenhagen, Denmark between 01.06.2007 and 01.06.2018. A CNN was trained on the retrieved data with the recorded treatment decisions as labels and validated on a subset of the data relabelled by three ophthalmologists to denote presence of oedema. Results: Moderate inter‐grader agreement on presence of oedema in the relabelled data was found (76.4%). Despite different training and validation labels, the CNN performed on par with inter‐grader agreement in detecting oedema on OCT images (AUC 0.97, accuracy 90.9%) and previously published models based on relabelled datasets. Conclusion: The level of performance shown by the current model might make it valuable in detecting disease activity in automated AMD patient follow‐up systems. Our approach demonstrates that high accuracy is not necessarily constrained by incongruent training and validation labels. These results might encourage the use of existing clinical databases for development of deep learning based algorithms without labour‐intensive preprocessing in the future.