Privacy-preserving federated learning for optic disc and cup segmentation in retinal fundus images

  • Forschungsthema:Medizinische Bildgebung für Modellierung und Simulation
  • Typ:Bachelor- oder Masterarbeit
  • Betreuung:

    M.Sc. Ciro Benito Raggio

  • Download als PDF

    Motivation
    Accurate segmentation of the major components of the retina (blood vessels, optic disc, and optic cup) is critical for diagnosing diseases such as diabetic retinopathy, hypertensive retinopathy, age-related macular degeneration, glaucoma[1].
     

    Centralized deep learning models have recently taken the lead in segmentation tasks that require large, aggregated datasets for high accuracy. However, this raises privacy concerns, especially for sensitive medical data such as retinal fundus images (Figure 1), and the difficulty of sharing such medical images has recently increased significantly.

    Can federated learning (FL) be used to develop a privacy-preserving method for optic disc and cup segmentation in retinal fundus images, that improves segmentation accuracy while preserving patient privacy and facilitating multi-institutional collaboration? The proposed approach involves multi-institutional collaboration to train robust models on different datasets without sharing sensitive data, thereby improving the overall performance and generalization of the segmentation methods.

     

    [1] Priyadharsini, R., et al. “Optic Disc and Cup Segmentation in Fundus Retinal Images Using Feature Detection and Morphological Techniques.” Current Science, vol. 115, no. 4, 2018, pp. 748–52.
    JSTOR.

    Student Project
    Your tasks include:
    1. Literature review
    2. Develop a federated learning framework for optic disc and cup segmentation in retinal fundus images using the provided code
    3. Evaluate the results obtained by comparing the metrics with results already available in the literature

    Skills needed:
    - Python, PyTorch framework
    - Deep learning fundamentals
    - Image processing
    - Medical images is plus
    - Federated learning is plus