Virtual histological staining of label-free imaging mass spectrometry using diffusion models

Introduction: : Imaging mass spectrometry (IMS) provides untargeted, multiplexed molecular mapping of biological tissues with chemical specificity and sensitivity; however, its intrinsic spatial resolution and morphological contrast remain inferior to those of conventional microscopy. To align IMS‐derived molecular signatures with precise cellular and tissue architectures, researchers typically undertake labor‐intensive histochemical staining, high‐resolution brightfield imaging, and complex multimodal image registration—procedures that are both time‐consuming and prone to alignment errors. Here, we introduce a diffusion model–based virtual staining (VS) strategy that digitally integrates histological detail directly into label‐free IMS images, enhancing the spatial resolution and histological contrast (Fig. 1A). Leveraging a Brownian bridge diffusion framework, our method synthesizes high‐fidelity VS images from IMS data possessing 10-fold larger pixel size, achieving cellular‐level morphology comparable to histochemically Periodic Acid–Schiff (PAS)–stained sections. By obviating the need for physical staining and complex multimodal image registration workflow, this technique offers significant benefits for mass spectrometry-based molecular histology and will help accelerate IMS-enabled histological analyses in life sciences.

Materials and

Methods: : The study assembled a dataset of label-free imaging mass spectrometry ion images derived from renal tissue sections of 14 patients, encompassing 1,453 selected mass-to-charge (m/z) channels at a spatial resolution of 10 μm per pixel. Corresponding high-resolution microscopic images (1 μm per pixel) were obtained by brightfield microscopy following PAS staining. Four tissue sections were allocated for model training, while the remaining ten were reserved for blind evaluation. The whole slide images used for training were cropped into smaller field-of-view (FOV) pairs and augmented to increase dataset diversity, resulting in a training set comprising 712 paired IMS–PAS image patches. Each IMS image measured 140×140 pixels, and its corresponding PAS-stained image measured 1400×1400 pixels. For blind testing, 201 paired IMS–PAS image patches were held out without augmentation, obtained from ten de-identified patients not included in the training set. Each testing pair consisted of an IMS image of 64×64 pixels and a corresponding PAS image of 640×640 pixels. Model training employed a Brownian Bridge Diffusion Model (BBDM), in which the forward diffusion process gradually corrupts the high-resolution PAS image with scheduled noise until it matches the low-resolution mass spectrometry image. During the forward diffusion process, a U-Net structure was trained to learn the reverse diffusion and denoise the noisy, low-resolution IMS-like intermediates back to the high-resolution PAS-stained image. During inference, the trained U-Net progressively denoised previously unseen IMS data through iterative reverse diffusion steps, producing VS images that closely match the morphological detail and contrast of the histochemically stained ground truth images.

Results, Conclusions, and Discussions:: After a one-time training effort, we validated our diffusion-based VS method on unseen renal tissue sections from the reserved 10 independent patients. Our model successfully transformed low-resolution, label-free ion images acquired by IMS—which possess 10-fold larger pixel size—into higher resolution virtually stained images matching the histochemically PAS stained brightfield microscopy images (Fig. 1B). Quantitative evaluation of YCbCr color distribution using the Hellinger distance across 201 VS–HS image pairs (Fig. 1C) demonstrated that the generated VS images achieved good color contrast and closely matched their corresponding histological references. Spatial frequency analysis of IMS, VS, and HS images for a representative field of view (Fig. 1D) further confirmed that our method substantially enhances the spatial resolution to a level comparable with conventional PAS staining. Importantly, assessments by a board-certified pathologist showed strong concordance in key histological structures—such as glomeruli (G), proximal tubules (P), and distal convoluted tubules (D)—between the virtually stained images and their histochemically stained counterparts (Fig. 1B), underscoring the clinical interpretability of the VS outputs. In conclusion, our diffusion-based virtual staining technique addresses the key limitations of IMS, notably its lower intrinsic spatial resolution and lack of cellular morphological detail. Additionally, our approach bypasses labor-intensive staining protocols, and the complex multimodal image co-registration typically required post-IMS acquisition. We anticipate that this technology will considerably advance molecular histology studies driven by imaging mass spectrometry, enhancing throughput and enabling more comprehensive tissue characterization in biomedical research.

Acknowledgements and/or References (Optional)::