Deep learning with attention mechanism for optical diffractive image reconstruction

Optical diffractive imaging needs to solve the inverse scattering problem to generate high-resolution quantitative images. Deep learning improves the solution to the electromagnetic inverse scattering problem for reconstructing the scattering potential. The proposed hybrid approach combines deep learning with wavelet transform as input features to a U-net architecture, allowing the model to focus on target areas of the scattered field. This method outperforms traditional iterative methods and improves further by integrating an attention mechanism into the input features. Our results on simulated datasets demonstrate the effectiveness of the proposed approach, with improved reconstruction accuracy upon integration of the attention module into the U-net architecture, achieving a structural similarity index score of 0.833 for the leukemia dataset and 0.9164 for the thalassemia dataset.