With applications in pathogen screening, cancer cell classification, and digital pathology, quantitative phase microscopy (QPM) quantifies optical phase differences to yield information about the morphology, dry mass, and dynamics of biological specimens. Being a label-free physical measurement, QPM is well suited to be integrated with modern physics-based machine-learning approaches. To this end, we present end-to-end differentiable physical and analytical neural models to build optimal QPM solutions for target applications. We further discuss the interpretability of our models from a statical viewpoint to critically evaluate their performance beyond simple accuracy tests.
|