Lens-free microscopy, while simple, faces challenges in reconstructing images from in-line holographic measurements, especially with high optical thickness samples. This paper introduces a novel approach alternating between inverse problem optimization and deep learning to overcome phase wrapping errors. A first reconstruction is fed into a neural network trained to correct these errors, and the network's output then initializes a second reconstruction step. This approach is demonstrated on cells in suspension at various densities, showing improved reconstruction quality compared to using inverse problem optimization alone.

L. Hervé, D. C. A. Kraemer, O. Cioni, O. Mandula, M. Menneteau, S. Morales, C. Allier