DL-DHM: A novel approach for out of distribution datasets in Deep Learning based off-axis Digital Holographic Microscopy

Karthik Goud Bujagouni, Swarupananda Pradhan,
DL-DHM: A novel approach for out of distribution datasets in Deep Learning based off-axis Digital Holographic Microscopy,
Optics & Laser Technology,
Volume 192, Part D,
2025,
113813,
ISSN 0030-3992,
https://doi.org/10.1016/j.optlastec.2025.113813.
(https://www.sciencedirect.com/science/article/pii/S0030399225014045)
Abstract: Generalizability to out of distribution datasets (OOD) is a bottleneck typically faced by Deep Learning based methods. In this work, we propose an approach to address out of distribution datasets encountered in deep learning based off-axis Digital Holographic Microscopy. First, a novel method for diverse dataset synthesis in off axis Digital Holographic Microscopy has been proposed for training Deep Learning networks. Secondly, a Deep Learning network DL-DHM has been custom designed by integrating two state of the art networks- Residual UNET and Vision Transformer. Experiments were conducted with custom built, off axis Digital Holographic Microscope on two classes of cells, namely Red Blood Cells and Epithelial Cheek Cells under different system parameters to evaluate the Deep Learning network’s performance in real world out of distribution data. Further research in the direction of advanced Deep Learning techniques and newer data generation methods will help improve the generalizability and performance of Deep Learning on diverse real world data encountered in holography.
Keywords: Digital Holographic Microscopy; Quantitative Phase Imaging; Deep Learning; UNET; Vision Transformer.