Transformer-based generative adversarial network enabled direct aberration determination

Structured illumination microscopy (SIM) has emerged as a powerful technique, surpassing the limitations imposed by optical diffraction and providing remarkable enhancements in both lateral and axial resolution compared with traditional diffraction-limited microscopy. However, it does come with certain limitations, including the need for a complex optical setup, extensive image acquisition, and computationally intensive post-processing. Motivated by the advancements in deep-learning-based super-resolution techniques, we propose an original three-dimensional (3D) representative learning algorithm called the transformer-based generative adversarial network (TransGAN), which can accurately predict corresponding aberrations through a combination of 17 mixed Zernike modes. Our approach outperforms state-of-the-art algorithms in various cellular structures, achieving impressive results with a mean square error of 2.358×10−5 for aberration determination. TransGAN presents a promising solution for enhancing SIM imaging, offering improved resolution and precise aberration estimation. This technique exhibits significant potential in overcoming the limitations associated with 3D SIM techniques and advancing the field of 3D optical microscopy.