DeepDOF-SE: affordable deep-learning microscopy platform for slide-free histology

Lingbo Jin, Yubo Tang, Jackson B. Coole, Melody T. Tan, Xuan Zhao, Hawraa Badaoui, Jacob T. Robinson, Michelle D. Williams, Nadarajah Vigneswaran, Ann M. Gillenwater, Rebecca Richards‐Kortum, Ashok VeeraraghavanNature Communications 2024 · DOI

Abstract

Histopathology plays a critical role in the diagnosis and surgical management of cancer. However, access to histopathology services, especially frozen section pathology during surgery, is limited in resource-constrained settings because preparing slides from resected tissue is time-consuming, labor-intensive, and requires expensive infrastructure. Here, we report a deep-learning-enabled microscope, named DeepDOF-SE, to rapidly scan intact tissue at cellular resolution without the need for physical sectioning. Three key features jointly make DeepDOF-SE practical. First, tissue specimens are stained directly with inexpensive vital fluorescent dyes and optically sectioned with ultra-violet excitation that localizes fluorescent emission to a thin surface layer. Second, a deep-learning algorithm extends the depth-of-field, allowing rapid acquisition of in-focus images from large areas of tissue even when the tissue surface is highly irregular. Finally, a semi-supervised generative adversarial network virtually stains DeepDOF-SE fluorescence images with hematoxylin-and-eosin appearance, facilitating image interpretation by pathologists without significant additional training. We developed the DeepDOF-SE platform using a data-driven approach and validated its performance by imaging surgical resections of suspected oral tumors. Our results show that DeepDOF-SE provides histological information of diagnostic importance, offering a rapid and affordable slide-free histology platform for intraoperative tumor margin assessment and in low-resource settings.

Introduction

In cancer surgery, the clinical gold standard for ensuring complete tumor removal is histopathology. Traditionally, intraoperative assessment relies on frozen sectioning, where resected tissue is frozen, sliced into micron-thin layers using a cryostat (typically costing ~$70,000), stained, and examined. This process is labor-intensive, takes 20–30 minutes, and requires specialized infrastructure often absent in low-resource settings.

Imaging fresh, unsectioned tissue presents two primary technical hurdles: subsurface scattering, which blurs images due to light penetration, and surface irregularity. Raw tissue is not flat, and standard high-resolution microscopes have a shallow depth-of-field (DOF), causing most of a "bumpy" sample to appear out of focus.

Methods

DeepDOF-SE (Deep Depth-of-Field Surface Excitation) addresses these challenges through a co-design of optical hardware and deep-learning software.

Optical Sectioning and EDOF

The system utilizes deep UV (280 nm) excitation, a technique known as Microscopy with Ultraviolet Surface Excitation (MUSE). UV light penetrates only a few microns into the tissue, effectively localizing fluorescence to the surface and eliminating background blur. To manage irregular surfaces, a physical phase mask is integrated into the optical path to encode depth information. A U-Net-based deep learning model then reconstructs these encoded images, achieving an extended depth-of-field (EDOF) of 200 µm—ten times greater than conventional objectives.

Virtual Histology and Training

Tissue is stained with inexpensive vital dyes (DAPI for nuclei, Rhodamine B for cytoplasm). To facilitate clinical interpretation, a semi-supervised CycleGAN translates these dual-channel fluorescent images into virtual Hematoxylin and Eosin (H&E) images. The EDOF network was trained on an eclectic dataset of 1,800 images, including oral cancer resections, TCGA histopathology, and natural scenes to ensure texture robustness. The CycleGAN was validated using 604 paired and unpaired images from human oral cavity resections and mouse tongue sections to prevent the "hallucination" of features.

Overview of the DeepDOF-SE platform, including the imaging pipeline (a), comparison with conventional fluorescence (b), and cost/infrastructure benefits (c).

Results

The DeepDOF-SE platform achieves a lateral resolution of 2.46 µm, sufficient for identifying subcellular features such as nuclear atypia. The 200 µm EDOF allows for rapid scanning of scalpel-cut tissue at 1.6 cm²/min without manual refocusing.

Validation on human oral tumor resections demonstrated high structural fidelity, with a multi-scale structural similarity index (MS-SSIM) exceeding 0.85 across the entire 200 µm depth range. In contrast, conventional microscopy performance dropped significantly (MS-SSIM ~0.39) when the sample was even slightly out of focus. Crucially, the system is highly cost-effective, with a total build cost under $7,000—approximately 1/20th the cost of a traditional histology suite.

Discussion

DeepDOF-SE offers a promising alternative for intraoperative margin assessment, particularly in resource-limited environments. By combining the physics of UV surface excitation with AI-driven focus and staining, the platform provides a rapid, "slide-free" workflow. While the system performs well on oral tissues, future work is needed to address specific challenges such as the appearance of adipose tissue, which differs from traditional H&E due to the absence of lipid wash-out in slide-free processing. Expanding the training datasets to include diverse organ types and conducting larger clinical trials will be essential for widespread clinical adoption.

Summary updated

Abstract

Histopathology plays a critical role in the diagnosis and surgical management of cancer. However, access to histopathology services, especially frozen section pathology during surgery, is limited in resource-constrained settings because preparing slides from resected tissue is time-consuming, labor-intensive, and requires expensive infrastructure. Here, we report a deep-learning-enabled microscope, named DeepDOF-SE, to rapidly scan intact tissue at cellular resolution without the need for physical sectioning. Three key features jointly make DeepDOF-SE practical. First, tissue specimens are stained directly with inexpensive vital fluorescent dyes and optically sectioned with ultra-violet excitation that localizes fluorescent emission to a thin surface layer. Second, a deep-learning algorithm extends the depth-of-field, allowing rapid acquisition of in-focus images from large areas of tissue even when the tissue surface is highly irregular. Finally, a semi-supervised generative adversarial network virtually stains DeepDOF-SE fluorescence images with hematoxylin-and-eosin appearance, facilitating image interpretation by pathologists without significant additional training. We developed the DeepDOF-SE platform using a data-driven approach and validated its performance by imaging surgical resections of suspected oral tumors. Our results show that DeepDOF-SE provides histological information of diagnostic importance, offering a rapid and affordable slide-free histology platform for intraoperative tumor margin assessment and in low-resource settings.
BibTeX
@inproceedings{jin2024deepdof,
  author = {Lingbo Jin and Yubo Tang and Jackson B. Coole and Melody T. Tan and Xuan Zhao and Hawraa Badaoui and Jacob T. Robinson and Michelle D. Williams and Nadarajah Vigneswaran and Ann M. Gillenwater and Rebecca Richards‐Kortum and Ashok Veeraraghavan},
  title = {{DeepDOF-SE: affordable deep-learning microscopy platform for slide-free histology}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-47065-2},
}
Figure