Deep Learning-Driven Virtual Staining and Birefringence Imaging of Amyloid in Unlabeled Tissue

Introduction: : Systemic amyloidosis comprises a group of protein-misfolding disorders in which insoluble fibrils accumulate in tissues, progressively impairing organ function. Definitive diagnosis requires Congo-red staining and manual inspection under cross-polarized light to reveal the apple-green birefringence of amyloid deposits. This workflow is hampered by hazardous reagents, labor-intensive slide handling, and reliance on specialized polarization microscopes—equipment not integrated into modern whole-slide imaging (WSI) scanners. Variability in stain quality, polarizer alignment, and subjective interpretation further introduce inter-observer inconsistency and diagnostic delays.
Recent advances in virtual staining via label-free autofluorescence imaging offer a promising alternative. Here, we present a deep-learning framework that transforms four-channel autofluorescence images of unstained cardiac biopsies into co-registered virtual brightfield and polarization views that closely mimic Congo-red–stained histology (Figure 1a). A conditional generative adversarial network (cGAN), augmented with an elastic registration module and guided by a pixel-wise modality selection map (DSM, digital staining matrix), is trained on paired autofluorescence–histology tiles; this trained model digitally generates virtual whole-slide images in minutes—without dyes, polarizers, or hardware modifications. This label-free, fully digital approach promises rapid, reproducible amyloid screening compatible with existing fluorescence-capable slide scanners.

Materials and

Methods: : Cardiac biopsy specimens from eight patients with confirmed amyloidosis were sectioned at 8 μm thickness. Unstained sections were imaged on an Olympus IX-83 fluorescence microscope (40×/0.95 NA) using DAPI, FITC, TxRed, and Cy5 filter sets, yielding four co-registered autofluorescence (AF) channels per field-of-view (FOV). Slides were then stained with Congo red, digitized in brightfield on a Zeiss AxioScan Z1 (20×/0.8 NA), and re-imaged under cross-polarized light on a modified IX-83 equipped with a polarizer–analyzer pair adjusted by a board-certified pathologist. We assembled 451 AF/brightfield/polarization tile triplets (2048×2048 pixels): 386 from six patients for training and validation (10% held out), and 65 from two patients for blind testing. Our cGAN comprises an attention U-Net generator, a patch-based discriminator, and a learnable elastic image registration head. A single-channel DSM (pixel value defined as: brightfield=+1, birefringence=–1) is concatenated to the AF inputs to direct the generator toward the desired modality. Models were trained on an NVIDIA RTX 3090 Ti using Adam (lr=2×10⁻⁵) until convergence. At test time, each 2048×2048 pixel label-free image tile is processed in < 2 s and stitched into a virtually stained WSI.

Results, Conclusions, and Discussions:: Our unified virtual staining framework closely replicates both Congo-red brightfield and apple-green birefringence from label-free AF data. Blind evaluations on 65 unseen test tiles (from two separate patients) demonstrated that virtual brightfield images reproduced congophilic pink regions and that virtual polarization images showed the characteristic apple-green signal (example FOVs shown in Fig. 1b). Three board-certified pathologists blindly scored histochemical and virtual images across seven metrics (M1–M7). Virtual slides matched or exceeded histochemical performance in 66–91% of FOVs for amyloid quantification, birefringence quality, and inter-modality consistency. Quantitative metrics— mean absolute error (MAE), multiscale structural similarity index measure (MS-SSIM), peak signal-to-noise ratio (PSNR), and Fréchet inception distance (FID), and down-sampled IoU—confirmed strong structural and color concordance.
By virtually generating both imaging modalities (brightfield and polarization) within a single DSM-driven network, we achieve perfect co-registration, avoid divergent artifacts, and eliminate hazardous chemicals, manual polarizer adjustments, and specialized optics. The inference pipeline processes a full WSI in minutes, offering a scalable, repeatable digital pathology solution. Because our method uses only standard autofluorescence filter sets, it can be integrated seamlessly into existing fluorescence-capable WSI scanners without hardware changes.
Future work will validate specificity against non-amyloid birefringent mimics, extend to other types of tissue (kidney, liver, spleen), and develop automated amyloid‐detection heatmaps. The DSM’s inherent multiplexing capability also enables additional virtual modalities—such as angle-shifted polarization imaging—to further bolster diagnostic confidence. Large-scale, multi-center trials will be essential to translate this fully digital, label-free staining paradigm into routine clinical practice.

Acknowledgements and/or References (Optional)::