An enhanced virtual phase contrast method for brightfield microscopy based on asymmetric illumination and deep learning acceleration

Brightfield microscopy remains a vital tool in biomedical research. However, when it comes to non-invasive imaging of unstained samples, its ability to provide clear visualization is limited. Although standard phase-contrast microscopy can address this issue, the need for matching annular stop and objective lens increases system complexity and cost. Additionally, the light restrictions imposed by the annular stop result in darker images, posing challenges for image acquisition. To overcome these challenges, this paper proposes an enhanced virtual phase contrast (VPC) method for brightfield microscopy based on asymmetric illumination and deep learning acceleration, combining brightfield and phase contrast microscopy advantages while mitigating their respective limitations. By utilizing a cylindrical lens to modulate the wavevector components of the illumination light, addressing the unreliability of directly generating virtual phase contrast images from brightfield images that lack phase information. Additionally, a data-driven conditional generative adversarial network (CGAN) with confidence maps is employed to accelerate the transformation of brightfield microscopy images into VPC images. Experimental results indicate that the reconstructed VPC images achieve image quality on par with conventional DPC images, and in certain scenarios, outperform them in terms of contrast and detail preservation. Furthermore, the proposed method demonstrates strong performance in applications such as cell counting and segmentation, providing an effective approach for enhancing brightfield microscopy images without requiring specialized phase contrast equipment.