Deep learning based virtual staining of label-free ultraviolet (UV) microscopy images for hematological analysis

Abstract

Deep ultraviolet microscopy (UV) enables high-resolution, label-free imaging of biological samples and yields diagnostically relevant quantitative molecular and structural information. We recently demonstrated that deep UV microscopy can serve as a simple, fast, and low-cost alternative to modern hematology analyzers that assess variations in the morphological, molecular, and cytogenetic properties of blood cells to monitor and diagnose blood disorders. We also introduced a pseudocolorization scheme that uses multi-spectral UV images (acquired at three different wavelengths) to generate images whose colors accurately recapitulate those produced by conventional Giemsa staining, and can thus be used for visual hematological analysis. Here, we present a deeplearning framework to virtually stain single-channel UV images acquired at 260 nm, providing a factor of three improvement in imaging speed without sacrificing accuracy. We train a generative adversarial network (GAN) using image pairs consisting of single-channel UV images of blood smears and their corresponding pseudocolorized images to generate realistic, virtually stained images. The virtual stained images are post-processed to improve contrast and yield consistent background colors. We quantify the performance of our framework in terms of the structural similarity index (SSIM) for each color channel. Our virtual staining scheme is the first step towards a completely automated hematological analysis pipeline that includes segmentation and classification of different blood cell types to compute metrics of diagnostic value. Our method eliminates the need to acquire images at different wavelengths and could potentially lead to the development of a faster and more compact label-free, point-of-care hematology analyzer.