PhaseStain: Digital staining of label-free quantitative phase microscopy images using deep learning (1807.07701v1)

Abstract: Using a deep neural network, we demonstrate a digital staining technique, which we term PhaseStain, to transform quantitative phase images (QPI) of labelfree tissue sections into images that are equivalent to brightfield microscopy images of the same samples that are histochemically stained. Through pairs of image data (QPI and the corresponding brightfield images, acquired after staining) we train a generative adversarial network (GAN) and demonstrate the effectiveness of this virtual staining approach using sections of human skin, kidney and liver tissue, matching the brightfield microscopy images of the same samples stained with Hematoxylin and Eosin, Jones' stain, and Masson's trichrome stain, respectively. This digital staining framework might further strengthen various uses of labelfree QPI techniques in pathology applications and biomedical research in general, by eliminating the need for chemical staining, reducing sample preparation related costs and saving time. Our results provide a powerful example of some of the unique opportunities created by data driven image transformations enabled by deep learning.

• The paper introduces a GAN-based framework that virtually stains label-free QPI images into conventional histochemically stained images.
• The method achieves high structural similarity scores (0.81 for skin, 0.81 for kidney, 0.89 for liver) while reducing time, cost, and tissue alteration compared to chemical staining.
• The approach offers robust, noise-resistant, and computationally efficient integration into digital pathology workflows, advancing precision diagnostics.

Overview of PhaseStain: Digital Staining of Label-Free Quantitative Phase Microscopy Using Deep Learning

The paper presents a novel approach to virtual staining of quantitative phase images (QPI) using a deep learning framework called PhaseStain. The core objective is to transform QPI images of label-free tissue sections into digitally equivalent images that would conventionally require histochemical stains like hematoxylin and eosin (H&E), Jones’ stain, or Masson’s trichrome. This transformation is achieved by utilizing a generative adversarial network (GAN) to map quantitative phase information to the amplitude contrast observed in stained brightfield microscopy images. The paper focuses on three tissue types: human skin, kidney, and liver.

Methodology and Implementation

The generative model was trained on paired QPI and traditionally stained brightfield images, allowing the GAN to learn the transformation from phase contrast to appearance contrast characteristic of stained samples. This methodology not only bypasses the need for traditional staining processes but also offers a significant reduction in time, cost, and potential tissue alteration through chemical staining. The deep neural network architecture employed in PhaseStain consists of two components: a generator that produces the virtual stains and a discriminator that distinguishes between virtual and authentic stained images to refine the generator's output. The networks were fine-tuned for different tissue types and staining protocols, demonstrating adaptability to various clinical scenarios.

Experimental Results and Quantitative Analysis

The PhaseStain framework was trained and tested on large datasets of holographically reconstructed images from different tissue types. The results showed that the virtually stained images demonstrated a high structural similarity index measure (SSIM), with values of 0.8113 for skin, 0.8141 for kidney, and 0.8905 for liver tissues, indicating strong visual resemblance to chemically stained counterparts. The framework was further evaluated for its robustness to noise and perturbations in the phase data, revealing a resilience that suggests utility in diverse pathological conditions. Moreover, the computational efficiency achieves fast inferencing speeds, enabling potential real-time integration into existing digital pathology workflows.

Implications and Future Directions

The PhaseStain methodology marks a significant advancement in label-free imaging technologies, extending the utility of QPI for clinical pathology and research. Since chemical staining is a time-consuming process prone to variability, PhaseStain offers a more consistent approach by standardizing the staining process through machine learning. This could augment current histopathological analysis, providing additional molecular insights through preserved tissue samples for downstream analyses, such as MALDI imaging or immunofluorescence tagging.

PhaseStain also highlights the growing role of artificial intelligence in medical imaging, particularly in replacing and augmenting traditional practices with data-driven methods. Potential future developments could entail the integration of more diverse stain types and tissue models, as well as improvements in the GAN’s architecture to further enhance the quality and applicability of the virtual staining process. Additionally, integrating PhaseStain with advanced imaging modalities beyond holographic setups could broaden the technique’s utility and accessibility in various medical and research settings.

By offering a digital alternative to histochemical staining, the PhaseStain paradigm may reshape workflows in pathology, not merely as an auxiliary tool but as a viable substitute that contributes to precision medicine, speeding up diagnostic processes and opening avenues for proactive patient management strategies.