Mathematical and Computational Engineering Stream

Robust and Generalizable Deep Learning for Retinal Image Analysis in Data-Constrained Settings

Faculty: Dr. Bhaskara Rao Chintada (BE) and Prof. Phaneendra K. Yalavarthy (CDS)

Proposal
Retinal imaging modalities such as Fundus Photography, Optical Coherence Tomography (OCT), and OCT-Angiography (OCTA) provide non-invasive diagnostic capabilities for several retinal diseases, enabling the diagnosis and monitoring of conditions including diabetic retinopathy, age-related macular degeneration, and glaucoma. While deep learning has demonstrated strong performance in retinal image analysis tasks like segmentation, vasculature extraction, and disease grading, its clinical translation remains hampered by poor generalizability. Models trained on data from a single imaging device or institution frequently experience performance degradation when applied to new settings. This challenge is further compounded by the high cost of expert annotation and the variability of real-world acquisition protocols across diverse patient populations.

This thesis work aims to develop robust, generalizable deep learning methods for automated retinal image analysis in data-constrained environments. A central focus will be physics-informed data augmentation, which introduces training-time variability reflective of the optical and physiological processes involved in retinal image formation. By modelling point-spread functions, imaging artifacts, speckle noise and correlation properties, and illumination non-uniformity, we aim to improve model robustness without requiring additional labelled data. Concurrently, the project will explore adapting large-scale vision foundation models (e.g., RETFound) to retinal image analysis tasks.

To further address the scarcity of labelled data and cross-site variability, the project will develop semi-supervised and weakly supervised learning frameworks, as well as domain generalization and test-time adaptation techniques for deployment across unseen devices and clinical sites.

Background Needed
Linear algebra, signal/image processing, computer vision, machine learning, programming.

Basic Qualifications
B.E./B.Tech. in EE/ECE/IN/CS/IT/DA/BME (or) M.Sc. (Mathematics/Physics)

References

  1. Zhou, Y., Chia, M. A., Wagner, S. K., Ayhan, M. S., Williamson, D. J., Struyven, R. R., … & Keane, P. A. (2023). A foundation model for generalizable disease detection from retinal images. Nature, 622(7981), 156-163.
  2. Ravishankar, H., Paluru, N., Sudhakar, P., & Yalavarthy, P. K. (2025, April). Inference Time Adaptation for Improved Retinal Disease Diagnosis Using Optical Coherence Tomography Images. In 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI) (pp. 1-5). IEEE.
  3. Chintada, B. R., Ruiz-Lopera, S., Bouma, B. E., Villiger, M., & Uribe-Patarroyo, N. (2025). Data augmentation strategies to improve the generalizability of deep learning-based OCT despeckling methods. In Optical Coherence Tomography and Coherence Domain Optical Methods in Biomedicine XXIX (p. PC133051C). SPIE.