AI for Ophthalmology

Posted on Wed 25 September 2024 in research

The application of Artificial Intelligence (AI) in ophthalmology has revolutionized the field by providing valuable support for medical decision-making. By leveraging machine learning algorithms and image analysis techniques, AI can aid clinicians in diagnosing and managing various eye conditions more accurately and efficiently. For instance, AI-powered systems can analyze retinal images from fundus photography, automated refraction, or optical coherence tomography (OCT) scans to detect subtle changes indicative of diseases such as diabetic retinopathy, age-related macular degeneration, or glaucoma. By identifying patterns and anomalies that may be missed by human observers, AI can enhance diagnostic accuracy, reduce false positives, and enable earlier interventions. Additionally, AI-driven predictive models can help identify patients at high risk of disease progression or complications, enabling personalized treatment plans and improved patient outcomes. Overall, the integration of AI in ophthalmology has the potential to transform clinical practice, improve patient care, and enhance the overall efficiency of eye care services.

Original image from the HRF dataset Image with vessels segmented

Example Application: Evaluation of a semantic segmentation techniques on color fundus imaging (CFI). True positives, false positives and false negatives are displayed in green, blue and red respectively.

Partnerships

  • Dr. Mattia Tommasoni, M.D. Florence Hoogewoud, Hôpital Ophthalmique Jules Gonin, Lausanne, Switzerland
  • Dr. Med. Christophe Chiquet, Department of Ophthalmology at University Hospital of Grenoble Alpes, France
  • Dr. Med. Christoph Amstutz, Augenklinik Luzerner Kantonsspital, Switzerland

Uveitis: Our recent work have made significant strides in developing automated grading systems for retinal inflammation. In (Amiot et al., 2023), we presented a novel pipeline capable of fully automating the grading of retinal vasculitis from fundus angiographies, achieving a high area-under-the-curve (AUC) score of 0.81, comparable to state-of-the-art approaches. Building on this work, in (Amiot et al., 2024), we developed an automatic Transformer-based grading system for multiple retinal inflammatory signs, utilizing a larger dataset of fluorescein angiography images. The model was trained on a dataset with 543 patients (1042 eyes, 40'987 images). The new approach demonstrated excellent performance in detecting vascular leakage, capillary leakage, macular edema, and optic disc hyperfluorescence. These advancements have the potential to streamline clinical evaluations, improve diagnostic accuracy, and enable more efficient patient care for this class of diseases.

Retinal vein occlusions: In (Mautuit et al., 2024) I contributed to the developement of a novel non-invasive tool, named AO-LDV, which combines adaptive optics with laser Doppler velocimetry to measure retinal venous blood flow in humans. This device enables accurate measurements of absolute blood flow rates and red blood cell velocities across various retinal vessel diameters. The study demonstrates that the AO-LDV can be used to quantify total retinal blood flow in healthy individuals (approximately 38 μl/min), which was found to correlate significantly with retinal vessel diameter and maximal velocity. The study's findings are also accompanied by a thorough evaluation of two software suites for automated retinal vessel measurement, one of which (based on deep neural networks) demonstrated higher accuracy and wider measurements.

Semantic segmentation: Our study (Galdran et al., 2022) presents significant advancements in retinal vessel segmentation from color fundus images, challenging the notion that increasingly complex deep learning models are necessary for high performance. By revisiting fundamental techniques and carefully training a minimalistic U-Net architecture, we demonstrated that it can closely approximate the performance of current state-of-the-art methods using orders of magnitude fewer parameters. Furthermore, they introduce a cascaded extension (W-Net) that achieves outstanding results on several popular datasets with even lower model complexity. The study also provides the most comprehensive cross-dataset performance analysis to date, highlighting the limitations of existing approaches and the potential for domain adaptation techniques. Overall, this work showcases efficient and effective solutions for retinal vessel segmentation tasks that align with current state-of-the-art results while reducing model complexity.

Demographic Fairness: Our study in (QueirozNeto et al., 2024) investigates the fairness and bias of Foundation models when applied to medical imaging datasets. By fine-tuning a Foundation model on the Brazilian Multilabel Ophthalmological Dataset (BRSET), researchers found that it had potential to reduce disparities in accuracy between different gender and age groups, compared to traditional supervised learning. However, as data availability decreased, the model's bias actually increased, suggesting that fairness issues may arise when deploying such models in real-world settings with limited data. These findings highlight the need to consider bias and fairness implications when using Foundation models in practical applications.


Bibliography

Victor Amiot, Oscar Jimenez-del-Toro, Pauline Eyraud, Yan Guex-Crosier, Ciara Bergin, André Anjos, Florence Hoogewoud, and Mattia Tomasoni. Fully automatic grading of retinal vasculitis on fluorescein angiography time-lapse from real-world data in clinical settings. In 2023 IEEE 36th International Symposium on Computer-Based Medical Systems (CBMS). June 2023. doi:10.1109/CBMS58004.2023.00301.

Victor Amiot, Oscar Jimenez-del-Toro, Yan Guex-Croisier, Muriel Ott, Teodora-Elena Bogaciu, Shalini Banerjee, Jeremy Howell, Christoph Amstutz, Christophe Chiquet, Ciara Bergin, Ilenia Meloni, Mattia Tomasoni, Florence Hoogewoud, and André Anjos. Automatic transformer-based grading of multiple retinal inflammatory signs on fluorescein angiography. September 2024. URL: https://papers.ssrn.com/abstract=4960069, doi:10.2139/ssrn.4960069.

Adrian Galdran, André Anjos, José Dolz, Hadi Chakor, Hervé Lombaert, and Ismail Ben Ayed. State-of-the-art retinal vessel segmentation with minimalistic models. Nature Scientific Reports, 12(1):6174, 4 2022. Number: 1 Publisher: Nature Publishing Group. URL: https://www.nature.com/articles/s41598-022-09675-y, doi:10.1038/s41598-022-09675-y.

Thibaud Mautuit, Pierre Cunnac, Frédéric Truffer, André Anjos, Rebecca Dufrane, Gilbert Ma\^ıtre, Martial Geiser, and Christophe Chiquet. Absolute retinal blood flow in healthy eyes and in eyes with retinal vein occlusion. Microvascular Research, 1 2024. doi:10.1016/j.mvr.2023.104648.

Dilermando Queiroz Neto, Anderson Carlos, Ma\'ıra Fatoretto, Luis Filipe Nakayama, André Anjos, and Lilian Berton. Does data-efficient generalization exacerbate bias in foundation models? In Proceedings of the 18th European Conference on Computer Vision (ECCV). October 2024.