Automatic quantification of Dry Eye Disease

Quantification automatique de la sècheresse oculaire

Automatic quantification of Dry Eye Disease

As our environment and lifestyle evolve, more people are suffering from dry eye disease. According to a recent study, this syndrome affects almost 7% of the American adult population [1]. Numerous tests are available to clinically quantify dry eye. According to the recommendations of The International Dry Eye WorkShop[2], the measurement of dry eye should be based on an examination of the ocular surface (cornea and sclera) using a slit lamp (or biomicroscope). The ophthalmologist first applies a fluorescent dye to the cornea, then observes several signs of dryness. Firstly, he measures the time it takes for the tear film to tear following a blink (BUT - Break Up Time), a short time indicating a thin or poor-quality film. Next, it detects areas of the ocular surface damaged by dryness (keratitis). These areas take the form of dots, spots or filaments: depending on their number and location, a degree of ocular dryness can be determined [3]. The problem with these measurements is that they are not very precise or reproducible. These limitations prevent reliable screening and quantification of ocular dryness in a patient.Our aim is therefore to implement an artificial intelligence system designed to perform these measurements accurately and reproducibly. This will improve longitudinal follow-up of patients and assessment of therapeutic response.

For several years now, Théa, a pharmaceutical laboratory specializing in research, development and marketing in ophthalmology, and the Laboratoire de Traitement de l'Information Médicale (LaTIM) - Inserm UMR 1110 have been working together to develop artificial intelligence algorithms to automate various tests. A Master 2 internship and a science thesis (https://theses.hal.science/tel-05006419) have already been carried out on this subject. In parallel, a database of videos of anterior segments has been acquired, enabling the algorithms to be easily trained.

Subject

The objectives of this thesis will be multiple and will focus on the automated analysis of videos by deep learning. The clinical objectives are the precise estimation of the “Oxford” score and the localization of tear film rupture. One of the approaches to be explored will be unsupervised super-resolution. Because of the video acquisition process, the same lesion will be observed in several frames. Exploiting this redundancy is essential to obtain good results.

This approach, based on unsupervised super-resolution, remains little explored in the medical field, although it has shown its effectiveness in areas such as satellite imagery [4]. More recent work uses generative AIs, notably variational autoencoders and GANs, to improve the quality of reconstructions [5][6]. Even more recently, diffusion models have emerged as a powerful alternative, due to their ability to generate high-quality images from random noise [7]. However, the combination of medical imaging, unsupervised super-resolution and diffusion remains a virgin area of research yet to be explored in the context of this thesis work. Following an in-depth bibliographical study, the first step will be to develop a high-performance super-resolution algorithm designed to improve the quality of video images and facilitate the segmentation of corneal lesions. On these improved images, an analysis model will be trained to automatically estimate the Oxford score. In collaboration with ophthalmologists, this score can be refined or enriched to offer the community a new, more precise and reproducible quantification scale. Super-resolved images will also be used to locate the initial zone of tear film rupture. This new clinical feature [8] is difficult to observe. In this case, the spatial and temporal redundancy of the videos will be an important lever in making detection more reliable.

Bibliographie/Bibliography

[1] K. F. Farrand, M. Fridman, I. Ö. Stillman, et D. A. Schaumberg, « Prevalence of Diagnosed Dry Eye Disease in the United States Among Adults Aged 18 Years and Older », Am. J. Ophthalmol., vol. 182, p. 90-98, oct. 2017, doi: 10.1016/j.ajo.2017.06.033.

[2] J. P. Craig et al., « TFOS DEWS II Definition and Classification Report », Ocul. Surf., vol. 15, no 3, p. 276-283, juill. 2017, doi: 10.1016/j.jtos.2017.05.008.

[3] J. P. Whitcher et al., « A simplified quantitative method for assessing keratoconjunctivitis sicca from the Sjögren’s Syndrome International Registry », Am. J. Ophthalmol., vol. 149, no 3, p. 405-415, mars 2010, doi: 10.1016/j.ajo.2009.09.013.

[4] K. Deepthi, A. K. Shastry, et E. Naresh, « A novel deep unsupervised approach for super-resolution of remote sensing hyperspectral image using gompertz-function convergence war accelerometric-optimization generative adversarial network (GF-CWAO-GAN) », Sci. Rep., vol. 14, no 1, p. 29853, déc. 2024, doi: 10.1038/s41598-024-81163-x.

[5] Z.-S. Liu, W.-C. Siu, L.-W. Wang, C.-T. Li, M.-P. Cani, et Y.-L. Chan, « Unsupervised Real Image Super-Resolution via Generative Variational AutoEncoder », in Conference on Computer Vision and Pattern Recognition, Seattle, Washington, United States: IEEE, juin 2020. Consulté le: 15 juillet 2025. [En ligne]. Disponible sur: https://hal.science/hal-02640889

[6] Z. Zhang, Y. Tian, J. Li, et Y. Xu, « Unsupervised Remote Sensing Image Super-Resolution Guided by Visible Images », Remote Sens., vol. 14, no 6, Art. no 6, janv. 2022, doi: 10.3390/rs14061513.

[7] H. Cao et al., « A Survey on Generative Diffusion Models », IEEE Trans. Knowl. Data Eng., vol. 36, no 7, p. 2814-2830, juill. 2024, doi: 10.1109/TKDE.2024.3361474.

[8] J. Kim, J. Y. Kim, K. Y. Seo, T.-I. Kim, H. S. Chin, et J. W. Jung, « Location and pattern of non-invasive keratographic tear film break-up according to dry eye disease subtypes », Acta Ophthalmol. (Copenh.), vol. 97, no 8, p. e1089-e1097, déc. 2019, doi: 10.1111/aos.14129.

As a public institution of a scientific, cultural and professional nature, the UBO’s primary missions are to contribute to the research development and to raise the scientific, cultural and professional level both nationally and internationally, to participate in regional and national growth, economic development and the implementation of employment and employability strategies. Finally, it contributes to reducing social and cultural inequalities.

- Master 2 or equivalent in Artificial Intelligence, image processing or biomedical computing

- Expertise in Python and deep learning frameworks (PyTorch)

- Skills in video processing, computer vision or self-supervised learning appreciated

- Interest in medical applications and working in collaboration with clinicians

- Fluency in written and spoken scientific English

- Good grades and references