3.1.

Limitations of Chemical Staining

One of the most impactful applications of DL for cross-modality transformation is in histology, where visual tissue analysis often faces challenges associated with traditional physical staining. Tissues, as the largest biological structures routinely observed through optical microscopy, require staining protocols to create visual contrast between features. However, these protocols often rely on chemical dyes that can be hazardous and may adversely affect the samples, especially during critical steps when sample structures are vulnerable.^25–27 In addition, manual labor and dexterity are necessary, and especially for fluorescence dyes, not all staining procedures are compatible in the same sample, limiting the information obtainable.

The histological process usually comprises several steps, including fixation, embedding, sectioning, staining, and mounting, although the specific steps may vary according to the staining technique and the target tissue. The first step is fixation, where the full tissue sample is preserved using chemicals, such as formaldehyde or glutaraldehyde. Fixation prevents decay and maintains the structural integrity by cross-linking the proteins in the sample. However, the tissue’s original chemistry is altered. An alternative approach is to freeze the sample, often using liquid nitrogen, which can preserve the natural state of proteins and lipids without chemicals. The next step is to dehydrate the tissue sample through a series of diluted alcohol solutions and to clear it, using different clearing agents that dissolve remnant lipids and simultaneously homogenize the refractive index. This process renders the tissue transparent as a consequence of even light scattering across the sample,²⁸ and prepares it for infiltration, but can cause tissue shrinkage and morphological alterations. The sample can now be completely encased in an embedding medium, such as paraffin wax, and left to solidify. Once hardened, the sample is cut using a microtome into very thin slices, around 4 to $10\mu \mathrm{m}$ thick, in a step known as sectioning. This process requires high skill and precision, as incorrect microtome alignment or use can lead to tearing or crushing of the tissue and obscure important details. Finally, the slices are placed on microscopic glass slides for observation. Handling must avoid stretching or folding, which can distort the samples and hinder analysis. Once mounted, the samples are ready for the next step: the staining. During staining, various stains are applied based on the cellular structures of interest. Among histological stains, hematoxylin and eosin (H&E) are among the most widely used, with hematoxylin staining cell nuclei purple and eosin coloring the cytoplasm and extracellular matrix pink. Other stains target different structures, such as Picrosirius Red (PSR) for collagen fibers or Alcian Blue for acidic mucins and cartilage. An important technique within this process is immunohistochemistry (IHC), which detects specific proteins in the sample using antibodies. This method requires antigen retrieval to unmask target proteins after fixation, alongside a blocking step to reduce nonspecific antibody binding. The staining process is sensitive to factors such as concentration and timing, and if applied unevenly, it can obscure details. Other limitations of IHC regard unspecific binding or cross-reactions of the added antibodies, which can negatively affect the outcome. Finally, the sample is prepared for detailed observation, and the histological features can be identified via microscopy imaging.

Traditionally a qualitative method, chemical staining can also yield quantitative data through image analysis,²⁹ such as measuring staining intensity, to estimate the presence of biomolecules. However, identifying biological features often requires additional context and expert analysis, and variability in staining intensity, reagent quality, and human interpretation can affect results. Standardized protocols and software tools can mitigate some of this variability, though issues persist compared to virtual staining and inherent contrast techniques.

As an alternative to chemical staining, inherent contrast techniques—such as phase contrast, differential interference contrast (DIC),^30,31 and quantitative phase imaging (QPI)—provide an alternative to chemical staining by exploiting refractive index variations in biological tissues to enhance contrast.³² These methods require minimal additional equipment but lack the specificity of chemical dyes and are prone to optical artifacts, such as halo effects.³³ To improve specificity, they are often combined with other techniques. The choice between virtual staining and inherent contrast methods depends on factors, such as cost-effectiveness and imaging quality.

3.2.

DL for Tissue Imaging

DL models can be trained to virtually stain samples, whether they are unstained³⁴ or have been stained using a different method.³⁵ By bypassing the physical staining process, multiple readouts equivalent to different dyes can be obtained from the same image. This not only maximizes information output³⁶ for analysis and diagnostics³⁷ but also simplifies the experimental setup requirements, as shown in Fig. 2.

Fig. 2

Contrast between physical and virtual approaches to obtain a stained image. In the physical approach, the sample undergoes a series of complex procedures, including preparation, staining, and imaging. Tissue preparation may involve fixing, embedding, and sectioning, among other steps. Similarly, histological staining of an unstained sample requires permeabilization, chemical dye application, washing, counterstaining, and protocol optimization before imaging. In contrast, virtual staining offers a simplified alternative to these protocols, eliminating the need for physical processing³⁷ or staining of the sample.³⁸ In the virtual approach, an unaltered or unstained sample is processed through a virtual staining network to generate a stained image, with results equivalent to physical staining. Physically stained images serve as training data, or input, for the model, especially when transforming between different stains is the objective. Created with the assistance of BioRender. (Tissue image adapted from Berkshire Community College Bioscience Image Library.)

In histology, DL models are trained to virtually stain tissue using collections of stain/unstained image pairs as, a reference,^{36,37,39–47} as shown in Fig. 3(a). This method constitutes a pure virtual approach to staining samples, although the starting point is not always an unstained sample. Some models have the capability to transform one type of staining into another.³⁶ Hong et al., as shown in Fig. 3(b),⁴⁸ washed and restained the samples to obtain equivalent pairs for a stain-to-stain translation model. Still, there are cases of cross-modality transformation of tissues based on unpaired samples.⁴⁹ These cases are notably more intricate as they typically do not rely on training pools consisting of paired samples,⁵⁰ but rather on disparate samples obtained through different techniques or dyes. In alternative scenarios, a multistain model,⁵¹ such as the one shown in Fig. 3(c), is developed, where the model is trained to virtually apply various dyes to an unstained sample, enabling it to transform a stained image into each of the other dye types.³⁶ An illustrative example is provided by Rivenson et al.,⁵² who used a CNN to transform wide-field autofluorescence images of unlabeled tissue sections⁵³ into images equivalent to the bright-field images of histologically stained versions of identical samples. Their study demonstrates the feasibility of this approach to generate multiple types of stains on different tissue types through the autofluorescence signal.

Fig. 3

Representative applications of cross-modality transformations for tissue imaging using DL. (a) Virtual staining of an unlabeled sample image to obtain the equivalent H&E stained image. Adapted from Rana et al.³⁹ (b) Stain-to-stain translation where the input and output are images from two different staining procedures, in this case H&E to IHC staining for cytokeratin (CK). Adapted from Hong et al.⁴⁸ (c) Multi-stain model that is able to transform unlabeled tissue images into different staining options simultaneously: H&E, orcein, and PSR. Adapted from Li et al.⁴⁰ (d) Cross-modality transform to apply a segmentation method, or potentially a stain, in a previously incompatible modality. In this case, an AI segmentation for MRI images is transcribed to CT images. Adapted from Dou et al.⁴⁹ (e) Biopsy-free cross modality transformation, where not only the staining procedure but also the sample preparation is avoided. Using CRM as a noninvasive technique for in vivo measurements, the resulting images incorporate features comparable to H&E, despite being incompatible with traditional staining techniques in such conditions. Adapted from Li et al.³⁷

Therefore, virtual and histological stainings are not mutually exclusive and can be complementary. While virtual staining offers convenience and reproducibility, it still relies on histological staining to provide the ground truth for generating large training data sets. The main trends are summed up in Table 1 in Sec. 6 guidelines. Some models benefit from existing databases of images of stained tissues for their training, reducing the manual effort required to obtain a sufficient training set.^34,61,62 Virtual staining also offers advantages over traditional histology, including the potential for real-time staining of tissue samples⁶³ and three-dimensional (3D) reconstructions of full tissues.⁶³ In the latter case, Wang et al.⁶³ virtually stained light-field microscopy—not to be confused with bright-field microscopy—images of volumetric samples, merging two typically incompatible techniques. Furthermore, certain virtual staining models extend their capabilities by integrating segmentation of the highlighted region of interest.^48,64–66 Segmentation, and potentially staining protocols, can be applied in imaging techniques where usually contrast-enhancing staining is not available due to cross-modality transformations, as shown in Fig. 3(d). In this example from Dou et al.,⁴⁹ segmentation, obtained through artificial intelligence in magnetic resonance imaging (MRI) measurements, is transferred to CT images. This is particularly relevant for the utilization of machine-learning models for diagnostic purposes, facilitating the differentiation of healthy tissue samples from those associated with conditions, such as infections or tumors.^61,62,67,68 Moreover, in certain cases,³⁷ not only diagnosis is accomplished through virtual staining, but also the direct acquisition of suitable input images from patients via noninvasive methods, as represented in Fig. 3(e), underscoring the potential of this powerful technique. While most cross-modality transformations in tissues are based on the same imaging modality with different nonfluorescent dyes, some examples involve transformation into fluorescent dyes³⁵ or other imaging techniques. Nevertheless, fluorescence is of greater relevance at the cellular level, where emphasis is placed on specific structures and organelles.

Virtual staining models must undergo rigorous validation against chemically stained samples to ensure accurate representation of biological features, as variations in color, texture, and detail may occur. A comprehensive comparison between traditional and virtual staining methods is therefore crucial for assessing reliability,⁶⁹ which requires either manual or automated ground-truth data annotation. Objective metrics, such as index measure (SSIM)^51,70–72 and peak signal-to-noise ratio (PSNR),^73–75 are used to measure accuracy, while expert visual evaluations are essential to confirm that virtual stains accurately reflect key histological features for research and diagnostic purposes.

In general, a single dye is insufficient to provide comprehensive information about a particular tissue sample. Instead, various dyes can be applied on different samples of the same original tissue, such as different slices of a single specimen. In the study by Li et al.,⁴⁰ three distinct dyes—H&E, PSR, and orcein—the last used to demonstrate elastic fibers—were utilized on carotid tissue. Each dye targets different components of artery tissue, aiding in the identification of coronary artery disease and vascular injuries. First, an independent model was trained for each dye to virtually stain an unstained sample. Then, a complete model was trained to simultaneously produce all three modalities from the original sample. This was accomplished by applying one of the three corresponding staining protocols to the originally unstained samples, generating pairs of stained and unstained images for each type. A total of 60 (whole slide images) of each stain, along with their unstained equivalents, were utilized, yielding 1500 to 1800 divided images for training and 150 to 200 images for validation for each staining protocol. Following standard practice, a conditional generative adversarial network (cGAN) was implemented to learn the generation of stained images from the acquired data set. The generator architecture was based on U-Net, while the discriminator comprised a PatchGAN architecture. To accomplish virtual staining according to three distinct protocols, the StarGAN⁷⁶ architecture was implemented. This framework enables image-to-image translations across multiple domains using only a single model, offering practically unlimited potential for utilizing unstained samples, as they can potentially be transformed into any other protocol with an appropriately trained network.^36,77

These studies suggest that DL holds significant potential for histological staining, yet its widespread adoption remains limited. Though DL has emerged as a leading choice for analyzing and interpreting histology images with the potential to enhance medical diagnostics,⁷⁸ very few algorithms have transitioned to clinical implementation.⁷⁹ Several challenges persist, notably the need for accurate labeling and addressing variations in slide colors,^80–82 as addressed in Sec. 6. Looking ahead, the availability of high-throughput experimental devices is crucial for optimizing the performance of DL-based digital histopathology methods. Fanous et al.⁸³ showed the potential of DL to speed up data acquisition in scanning microscopes, using a GAN-based image restoration approach to reconstruct motion-blurred scanned images.⁸⁴ This suggests that DL-based approaches could significantly improve experimentation efficiency and speed, thereby improving algorithm performance. In addition, integrating multiple modalities, such as molecular profiling information,^85,86 could further enhance the accuracy of disease classification and prognosis accuracy in digital pathology. Furthermore, the development of transfer-learning techniques that profit pretrained models on large data sets could mitigate the challenge of limited training data in certain biological applications, expanding DL’s utility in digital pathology. Promising advancements are expected in the use of diffusion models for data augmentations and synthetic image generation, as demonstrated by Moghadam et al.,⁸⁷ where the generated histopathology images were indistinguishable by experienced pathologists.

Diffusion models have emerged as a powerful tool in histology for virtual staining, addressing some of the limitations seen in traditional models, such as GANs and encoder–decoder networks. For instance, StainDiff is a diffusion probabilistic model designed to improve the stain-to-stain transformations by overcoming issues, such as mode collapse, where the generator of a GAN produces images based on a limited range of the training samples and posterior mismatching found in other networks.⁸⁸ These models are also applied to generate virtual IHC images from H&E stained slides, as seen in PST-Diff, which ensures structural and pathological consistency through mechanisms such as asymmetric attention and latent transfer.⁸⁹ Despite their potential, diffusion models generally require large data sets, making them less feasible for histological applications with limited data. To address this, multitask architectures such as StainDiffuser have been developed to simultaneously generate cell-specific stains and segment cells, optimizing the performance even with constrained data sets.⁹⁰ In addition, advanced methods, such as virtual IHC multiplex staining, utilize large vision-language diffusion models to generate multiple IHC stains from a single H&E image, addressing tissue preservation challenges often faced in biopsies.⁹¹ However, challenges remain. Diffusion models have shown limitations in unpaired image translation tasks, such as slide-free microscopy virtual staining, where the sample preparation process is bypassed along with the staining process. In such cases, they underperform compared to models such as CycleGAN without additional regularization, highlighting the need for further refinement in certain applications.⁹² Despite these challenges, diffusion models show great promise in virtual staining, with ongoing research focused on enhancing their reliability and applicability in histology.

4.1.

Limitations of Fluorescence Staining

Standard cell imaging workflows typically rely on fluorescence microscopy, employing either fixed or live fluorescent staining techniques to highlight specific cell structures. Despite their widespread use, both fixed and live fluorescent staining methods have limitations. These procedures can be invasive and toxic, potentially impacting cell health and behavior.⁹⁷ In fixed staining, as for tissue, the fixation process itself can introduce artifacts by altering the native state of cellular components. In addition, the use of permeabilization methods compromises cell membrane integrity.⁹⁸ Furthermore, fixed staining provides only a static view of cellular processes, limiting the ability to study dynamic processes. Conversely, live staining, while theoretically preserving the native state of cells, often alters their biological activity and can be toxic.⁹⁹ The availability of specific and effective live-staining dyes can also be limiting, restricting the visualization of certain cellular components. In addition, real-time observation of cellular processes in live staining may be challenging due to phototoxicity and photobleaching over extended imaging periods.¹⁰⁰ Lastly, the use of multiple fluorophores can lead to spectral cross talk between fluorescence channels, potentially resulting in misleading results and complicating image analysis. These challenges hinder the acquisition of reliable longitudinal data, which is often crucial for studying the effects of drug exposure over time.¹⁰¹

4.2.

DL for Cellular and Subcellular Imaging

Recently, research has proposed the use of DL as an alternative to conventional physical staining methods to mitigate inherent problems. These works suggest replacing physical staining and fluorescence microscopy with a neural network that generates virtual fluorescence-stained images from unlabeled samples.

Virtual cell staining has been achieved from various imaging modalities, including phase contrast,^55,102 QPI,¹⁰³ and holographic microscopy.¹⁰⁴ Moreover, recent studies have shown that bright-field images, despite their limited detail, contain sufficient information for a CNN to reproduce different types of staining.

For example, Ounkomol et al.¹⁰⁵ introduced a CNN-based framework to map the relationship between paired 3D bright-field and fluorescence live-cell images for various key subcellular structures (e.g., DNA, cell membrane, nuclear envelope, and mitochondria). Each cellular component is modeled separately, with a U-Net trained independently for each one. The training process minimizes the mean-squared error between the ground-truth fluorescence image and the predicted image. Once trained, these models can be combined, allowing a single 3D bright-field input to generate multichannel, integrated fluorescence images across multiple subcellular structures. Particularly advantageous, the training data require relatively few paired examples (solely 30 pairs per structure), lowering the machine-learning entry barriers.

The work by Helgadottir et al.⁵⁴ offers yet another compelling example of the potential of virtual staining of cellular structures from bright-field images. Similar to other approaches, this method relies on a modified version of the U-Net to learn the cross-modality mapping. However, it enhances the reconstruction accuracy by incorporating GAN-based training.

GANs have become a widely adopted framework in virtual cell staining due to their capacity to generate high-quality, realistic images.^{54,55,102–104} Particularly, Helgadottir et al. employed a conditional GAN to virtually stain lipid droplets, cytoplasm, and nuclei from bright-field images of human stem-cell-derived adipocytes. The generator network, a U-Net, processes a stack of bright-field images captured at multiple $z$ positions and generates virtually stained fluorescence images. The architecture features three independent decoding paths, each dedicated to one cellular component, to effectively decorrelate the features associated with the predicted fluorescence images. The discriminator, a CNN, is trained to distinguish between the synthetically generated virtual stains and actual fluorescently stained samples (conditioned on the input bright-field image). These two neural networks are trained simultaneously until the generator can fool the discriminator by producing images that closely mimic real fluorescence [Fig. 4(a)]. An interesting feature of Helgadottir’s method is its robustness and rapid convergence; the neural network requires relatively few epochs to quantitatively reproduce the corresponding cell structures [Fig. 4(b)].

Fig. 4

Virtual cell staining using DL. (a) Helgadottir et al. introduced a cGAN to virtually stain lipid droplets, cytoplasm, and nuclei using bright-field images of human stem-cell-derived fat cells (adipocytes). The U-Net-based generator processes bright-field image stacks captured at various $z$ positions to generate virtually stained fluorescence images. A CNN-based discriminator is trained to differentiate between the virtually generated stains and real fluorescently stained samples, conditioned on the input bright-field image. (b) The virtual staining of lipid droplets (green channel) and cytoplasm (red channel) exhibits a high degree of fidelity, as evidenced in the fine details of the lipid droplet internal structure and the enhanced contrast among distinct cytoplasmic components (highlighted by the arrows). Panels (a) and (b) adapted from Helgadottir et al.⁵⁴ (c) Unsupervised cross-modality image transformation using UTOM. Two GANs, G and F, are trained concurrently to learn bidirectional mappings between image modalities. The model incorporates a cycle-consistency loss ($L_{\text{cycle}}$) to ensure the invertibility of transformations, while a saliency constraint ($L_{\mathrm{sc}}$) preserves key image features and content in the generated outputs. (d) UTOM achieves performance comparable to a supervised CNN trained on paired samples, without requiring paired training data. Panels (c) and (d) adapted from Li et al.⁵⁵ (e) Co-registered label-free reflectance and fluorescence images acquired using a multimodal LED array reflectance microscope. (f) Multiplexed prediction displaying DNA (blue), endosomes (red), the Golgi apparatus (yellow), and actin (green). Zoomed-in views, with white circles, highlight representative cell morphology across different phases of the cell cycle. Adapted from Cheng et al.¹⁰⁶

While GANs have proven effective in enhancing the performance of virtual staining networks for various applications, they rely on co-registered input and ground-truth images. Nevertheless, obtaining perfectly co-registered training pairs is often challenging due to the rapid dynamics of biological processes or the incompatibility of different imaging modalities. To address this limitation, Li et al.⁵⁵ introduced unsupervised content-preserving transformation for optical microscopy (UTOM). This approach utilizes a CycleGAN to transform images between domains without requiring paired data. Unlike traditional GAN models, CycleGANs employ two generator-discriminator pairs, one for each domain, to learn bidirectional mappings between imaging modalities [Fig. 4(c)]. UTOM has been applied, among other examples, to the virtual staining of phase-contrast images of differentiated human motor neurons, notably delivering competitive performance compared to a CNN architecture trained on paired samples under supervision despite the lack of paired training data [Fig. 4(d)].

Importantly, although the architecture and training of the neural network play a decisive role in the performance of virtual staining models, the input imaging modality must capture sufficient contrast of the different cell structures, providing the network with enough information to learn the transformation to the desired high-contrast, high-specificity fluorescently stained samples.

Recent research has centered on the development of optical systems that capture the rich structural details of cells and embed inductive bias within the network to enhance its performance. For instance, Cheng et al.¹⁰⁶ profited from the rich structural information and high sensitivity in reflectance microscopy to boost the performance of virtual staining models. Specifically, the authors employed an LED array reflectance microscope to acquire co-registered label-free reflectance and fluorescence images.¹⁰⁷ This platform collects four dark-field reflectance images using half-annulus LED patterns oriented in different directions (top, bottom, left, and right). These measurements derive two dark-field reflectance differential phase contrast (drDPC) images computed along orthogonal orientations [Fig. 4(e)]. Interestingly, the oblique illumination dark-field and drDPC images provide complementary contrast information. While raw dark-field images highlight subcellular structures, such as nuclei, nucleoli, and hyperreflective areas near the nuclear periphery, the drDPC images emphasize cell membranes with clearly defined boundaries. These images serve as multichannel input for the virtual staining model, which, boosted by the enhanced resolution and sensitivity in the backscattering data, provide a reliable prediction of subcellular features [Fig. 4(f)].

In a similar vein, Cooke et al.¹⁰⁸ proposed incorporating a physical model of the experimental microscope into the virtual staining model. This approach utilizes a CNN which incorporates a “physical layer” representing the microscope’s illumination model. Consequently, during training, the network learns task-specific LED patterns that significantly enhance its ability to infer fluorescence image information from label-free transmission microscopy images. This work, in particular, further underscores the importance of rich input data and highlights the potential combination of programmable optical elements and physics-informed DL to open new possibilities for exploring the structure and function of cells.

5.1.

Physics of Superresolution Microscopy

When light from a point-like light source (an object with diameter $D_1$) traverses a lens, it undergoes diffraction, producing a characteristic pattern known as the Airy disk. This pattern comprises a bright central region surrounded by concentric rings of diminishing intensity ($D_2$). The Airy disk represents the smallest focal point achievable by a light beam. Below the object and Airy disk representations, corresponding intensity plots illustrating the point spread functions (PSFs) are displayed. As Airy patterns reach a point of significant interference, causing a reduction in contrast, they merge, becoming indistinguishable and limiting the spatial resolution. Spatial resolution, the shortest physical distance between two points within an image, stands out as the single most important feature in optical microscopy.¹¹³ The primary constraints affecting the achievable spatial resolution stem from an intrinsic phenomenon of diffraction physics. Regardless of lens quality or optical component alignment, a microscope’s resolution ultimately correlates with the wavelength of the detected scattered light and inversely with the numerical aperture (NA) of its objective. This relationship is shown in Fig. 5(a), where light from the sample traverses an objective to the image plane, generating a fundamentally limited diffraction pattern known as a PSF. The PSF inherently limits the minimal distance between two discernible points in the sample, shown in Fig. 5(b). The full width at half-maximum of a PSF in the lateral directions can be approximated as $\mathrm{FWHM}=0.61\frac{\lambda }{\mathrm{NA}}$, where $\lambda$ represents the light’s wavelength, and NA denotes the numerical aperture of the objective. Thus, for a typical oil immersion objective with $\mathrm{NA}=1.4$, the resulting PSF has a lateral size of 200 nm and an axial size of 500 nm, effectively restricting the resolution to this range for visible-light studies.¹¹⁴ Comparing these scales with those depicted in Fig. 1, it becomes evident that the diffraction limit rarely poses a challenge in most imaging at organ, tissue, or even cellular levels. However, in cellular exploration, where subcellular and molecular structures are of interest, issues regarding diffraction limits become prominent. These issues are exacerbated by the typically dense distribution of molecules and subcellular structures, causing their PSFs to overlap, thus blurring many intricate details together. Hence, the development of superresolution techniques that surpass the diffraction limit becomes imperative for further exploration of these structures using noninvasive optical light. Various microscope techniques have been developed to overcome this limitation, including single-molecule localization microscopy (SMLM) methods, such as STORM,¹¹⁵ photo-activated localization microscopy (PALM),¹¹⁶ and fluorescence photoactivation localization microscopy.¹¹⁷ Other methods of transcending the standard resolutions of microscopes exist, including complex numerical estimations of point spread (transfer) functions seeking to estimate the diffraction behavior, illumination pattern engineering methods reducing the PSF size,¹¹⁸ as well as specialized fluorophores.¹⁰⁹ However, these approaches pose their own challenges, including complex and multivariate dependencies on imaging conditions, making solving diffraction integrals of PSFs exceedingly difficult for practically relevant systems,¹¹⁹ as well as increased costs associated with the aforementioned fluorophores.¹²⁰

Fig. 5

Superresolution physical principles. (a) Illustration of the PSF resulting from imaging object of diameter $D_1$ below the diffraction limit of an optical system, leading to an image of diameter $D_2$. (b) Low-resolution image of simulated emitters alongside ground-truth emitter positions. Image reproduced with permission from Ref. 56.

In recent years, another promising avenue for achieving super-resolution has emerged as a consequence of the astounding growth and success of DL-based computer vision algorithms. Analogous to the cross-modality transforms mentioned above, the DL approach to superresolution involves training neural networks to transform one imaging modality (regular-resolution images) to another (superresolved images). Some of these approaches utilize generative learning, effectively learning the complex interpolation function between regular- and superresolved images, or through direct supervised learning, for example, by estimating the positions of underlying diffraction-limited emitters. The specific techniques for training these networks vary considerably across applications, as elaborated upon further below.

5.2.

DL for Superresolution Microscopy

In general, DL for superresolution can be categorized into two approaches, each with two learning paradigms. The first approach aims to enhance resolution by training end-to-end, directly transforming low-resolution images into high-resolution ones. This can be achieved through supervised learning, using pairs of simulated or experimentally measured images from the same sample at different resolutions to train neural networks, or through unsupervised learning, where only low-resolution or high-resolution images are obtained, either experimentally or through simulations. The other approach seeks resolution enhancement by training a network to output the position of each individual molecule (or equivalent scattering object) within an image, and then reconstructing the high-resolution image from these positions, thus transcending the diffraction limits. This approach can also be trained either in a supervised or unsupervised fashion. A summary of the different models and their characteristics can be found in Table 1 in Sec. 6 guidelines. Once such a network is trained, it can swiftly generate high-resolution images without the need for parameter adjustment, yielding an efficient algorithm for improving image resolution within a specific modality.^121–126

5.2.2.

Specific architectures and methods

Although the generic approaches mentioned above can result in a practically infinite variety of specific architectures, many significant superresolution studies in molecular microscopy have been achieved using a small number of named architectures. Structured feature superresolution microscopy model architecture, shown in panels of Fig. 6 as an example, allows for precise live-cell imaging with high spatial and temporal resolution to continuously monitor subcellular dynamics over extended periods. Among these, ANNA-PALM,⁶⁰ a U-Net based cGAN trained solely with experimental data, stands out; Deep-STORM,⁵⁶ based on a CNN encoder–decoder network trained with simulated data; and smNet,⁵⁷ which directly outputs molecule location, dipole orientation, and wavefront distortion from complex and subtle features of the PSF. In single-molecule superresolution microscopy, there is generally a trade-off between throughput and resolution. To construct a high-quality superresolution image, a large number of molecules need to be localized with high precision, requiring sufficient localizations before sampling a structure of interest.

Fig. 6

Superresolution applied architecture. The superresolution network enhances image resolution by training on pairs of simulated low-resolution (LR) and high-resolution ground-truth images or on wide-field (WF) and STORM images from a STORM microscope. First, the LR/WF image undergoes preprocessing through a subpixel edge detector to generate an edge map, both of which serve as inputs to the network. Training is guided by a multi-component loss function that incorporates the combination of multiscale structure similarity index measure and mean absolute error loss (MS-SSIM L1) to capture pixel-level accuracy between the superresolution (SR) and ground-truth/STORM images through multiscale similarity and mean absolute error, perceptual loss to assess feature map differences via the visual geometry group network, adversarial loss using a U-Net discriminator to differentiate ground-truth/STORM images from SR images, and frequency loss to compare differences in the frequency spectrum between SR and ground-truth/STORM images within a specific frequency range using the fast Fourier transform function. This comprehensive loss function helps the superresolution network model achieve precise and perceptually accurate superresolution imaging. Image adapted from Chen et al.¹⁴⁷

ANNA-PALM accomplishes this by training on a set of blinking single molecules from which a high-quality superresolution image can be experimentally acquired. A subset of these frames is used to generate a (low-quality) “sparse” superresolution image, which, alongside the diffraction-limited image and information about the imaged structure, serves as inputs into ANNA-PALM. The output, or label, is the full superresolution image reconstructed using all frames. Once trained, ANNA-PALM demonstrated the ability to provide high-quality results in imaging mitochondria, the nuclear core complex, and microtubules⁶⁰ at significantly higher speeds than conventional methods.

ANNA-PALM has proven to be a valuable method for accelerating the acquisition of high-density superresolution images by several orders of magnitude. However, there are many other DL models more appropriate for direct single-molecule localization. An important early model for this is Deep-STORM, developed for the acquisition of superresolution images of microtubules with single or multiple overlapping PSFs. While non-DL algorithms exist for this purpose,¹⁴⁸ they typically suffer from high computational costs and require sample-specific parameter tuning. Deep-STORM, an encoder–decoder CNN trained on simulated images, consists of simulated PSFs in various positions in an image on top of experimentally relevant background levels. Said simulated images are thereafter upsampled by a constant factor, constituting a superresolution image. These two versions of the same image are fed as input and output, respectively, to Deep-STORM during training. Such a model has been used to superresolve images of microtubules and quantum dots,⁵⁶ and inspired works on localizing high-density ultrasound scatterers¹⁴⁹ and Crispr-CAS-protein-DNA binding events.¹⁵⁰

Another impactful model is the aforementioned smNet, which similarly employs a (ResNet-inspired) CNN trained on simulated images for superresolution. The key distinctions lie in the image recreation of 3D images and the network’s outputs of the 3D coordinates and orientations of PSF-convoluted emitters from which the superresolution image can be reconstructed. smNet has been demonstrated to localize highly astigmatic single-molecule PSFs in experimental images of significantly higher quality compared to conventional Gaussian fitting methods.⁵⁷ This approach, reminiscent of DeepLoco,¹⁵¹ was developed around the same time and trains NNs to reconstruct simulated emitters in 3D through well-defined mathematical models of astigmatic PSFs. Other related examples of 3D superresolution are that of Zhou et al., who used a dual-GAN framework to directly superresolve images of mouse brains and bodies taken with fluorescence microscopy,¹⁵² or Zhang et al.¹⁵³ and Zelger et al.,¹⁵⁴ who used a U-Net-based and CNN-based approach, respectively, for superresolution in SMLM.

For images with higher PSF density, Speiser et al. introduced the method known as DECODE.⁵⁸ This architecture consists of a stack of two U-Nets, where the first U-Net processes a feature representation of a single frame, and the second U-Net processes feature representations of consecutive frames. The output of this method consists of several channels, each containing information in each pixel of the input image regarding (1) the probability of containing an emitter, (2) its brightness, (3) its 3D coordinates, (4) its background intensity, and (5) epistemic uncertainty of its localization and brightness. In the work of Speiser et al.,⁵⁸ this architecture is trained on simulated PSFs with a loss function connected to all five aforementioned types of pixel-level information and has been successfully applied to microtubules in conditions of low light exposure and ultrahigh sample densities.

6.1.

Data Quality, Augmentations, and Data Normalization

The quality of data plays a critical role in determining the performance of DL models. Two major issues commonly affect model quality: insufficient data to capture the variability within the data set or training data that fail to represent the conditions under which the model will be applied.

To detect the issue of insufficient data, a standard approach is to set aside a validation set that the model never sees during training. If the model’s performance on this validation set is significantly worse than on the training set, it likely indicates a lack of sufficient training data to generalize effectively. A common practice is to allocate ∼20% to 30% of the data set as a validation set. However, it is important to ensure that the validation set is maximally decorrelated from the training set to avoid misleading results. For example, it is ideal to sample from different locations in the sample or even from entirely different experimental videos. A poor sampling strategy, such as selecting every fifth frame from the same video, would introduce a high correlation between the training and validation sets, resulting in overly optimistic performance estimates.

If detected, data scarcity can be mitigated by data augmentation techniques to synthetically increase the diversity of training data. Techniques such as geometric transformations (rotation, scaling), noise injection, and intensity variation can simulate a broader range of conditions. However, care must be taken to ensure that these transformations can be meaningfully applied across modalities. For instance, intensity variations in quantitative phase contrast imaging hold physical significance and altering them synthetically could distort biologically relevant information. Geometric translations, in most cases, provide limited benefit, as convolutional models are inherently translation-equivariant. However, if the chosen model breaks translation equivariance (such as vision transformers), they may be useful.

Regularization techniques also play an essential role in improving model robustness, especially when data are scarce or noisy. Methods, such as dropout, weight decay, or $L_2$ regularization, are commonly used to prevent overfitting by penalizing overly complex models that may fit noise in the data rather than in underlying patterns. In scenarios where the model could easily memorize the training data, these techniques ensure that the model learns generalizable features rather than artifacts specific to the training set. Advanced regularization techniques, such as Bayesian regularization, can further improve robustness by incorporating uncertainty into the model’s predictions, making it especially useful in tasks where noisy or variable data are expected.

Transfer learning offers another potential solution to address limited data availability. Pretrained models, especially those trained on large data sets from similar domains, can be fine-tuned to perform specific tasks in microscopy. By leveraging features learned from related tasks, transfer learning reduces the need for extensive training data while still allowing the model to generalize effectively. This approach not only speeds up training but also improves the model’s performance on smaller, domain-specific data sets. In some cases, transfer learning from pretrained models in related fields, such as medical imaging, can be more effective than starting from scratch, especially in scenarios where biological structures share visual characteristics across different imaging modalities.

When the training data are nonrepresentative, this issue can be identified by observing a drop in model performance under real-world conditions, even though the performance on the validation set remains strong. This discrepancy often arises due to variations in optical systems, sample preparation protocols, or environmental factors that differ from those present in the training data. For instance, subtle differences in microscope settings, sample staining techniques, or even temperature can cause a shift in the data distribution, leading to poor generalization when the model is applied in different scenarios.

The primary strategy to address issues of representativeness is through effective data normalization. Normalization techniques aim to reduce variability in the data by standardizing features across data sets, such as intensity scaling, contrast adjustment, or color normalization. This can help minimize discrepancies between data sets generated under different conditions. However, caution must be taken in modalities where quantitative relationships between intensity values are critical. In such cases, aggressive normalization may disrupt important mappings between intensity and biological features, potentially degrading the model’s ability to learn meaningful cross-modality transformations. Furthermore, domain adaptation techniques can be employed to align the distributions of training and application data, improving the robustness of the model across diverse conditions.

6.2.

Model Selection

The choice of model architecture can have a significant impact on the performance of the model and depends on several key factors, such as data availability, target task, and specific requirements. Here, we give a general guideline for choosing an appropriate model.

If your data are not aligned, the CycleGAN is recommended. Aligned data refer to cases where each image in one modality has a direct counterpart in the other modality, meaning that both images capture the same sample part of the sample under the same conditions, making it possible to map pixel-to-pixel relationships between the two. When such paired data are unavailable, CycleGAN is suitable because it learns to map between modalities without requiring this strict correspondence. However, the less restrained training procedure also is likely to result in the model learning transformations that are less precise or biologically relevant, especially when precise quantitative relationships between modalities are required. Careful evaluation and additional constraints may be necessary to ensure that the model’s outputs are meaningful and accurate.

Assuming your data are paired and aligned, we recommend starting with a conditional GAN architecture, specifically using a U-Net-like generator and a spatial discriminator. A well-established configuration for this setup is the pix2pix model. Conditional GANs are optimized to generate quantitative, physically meaningful images by leveraging paired data to learn a direct mapping between input and output modalities. The U-Net generator was originally developed for biomedical images and is one of the most proven and widely adopted architectures for tasks involving fine-scale structural details. Spatial discriminators, in turn, evaluate the realism of local regions of the image rather than assessing it as a whole, often resulting in more detailed and accurate outputs.

However, depending on the specific requirements of your task, other architectures may be better suited. For example, if the goal is simply to enhance the contrast of specific substructures without requiring physical realism in the produced images, it may be more practical to forego generative models entirely. In such cases, direct supervised training of a U-Net can offer a simpler and stabler solution. The drawback of using a purely supervised U-Net is that it may lack the ability to generate the nuanced, high-fidelity details that generative models, particularly GANs, are capable of producing. However, for applications where interpretability and stability are more important than photorealism, this trade-off can be worthwhile.

On the other hand, if maximal photorealism is required, diffusion models are worth considering. These models have consistently been shown to produce highly realistic images, often outperforming GANs in terms of image quality and stability. Diffusion models work by iteratively denoising random noise to generate an image, which allows them to better capture fine-grained details and complex textures. However, diffusion models are typically much more computationally expensive, both to train and to evaluate, compared to GANs. Moreover, one should be careful not to conflate photorealism with better quantitative performance on downstream tasks.

Another important consideration is the spatial distribution of information in the image. The U-Net generator is highly effective for analyzing local, position-invariant features, making it ideal for tasks where the meaning of a structure does not depend on its specific location within the image. However, for data where the spatial context is crucial, such as brain scans, attention-based models may be more suitable. Attention mechanisms allow the model to focus on specific regions of the image while considering their global relationships, enabling more context-aware analysis. This makes attention-based architectures a better choice for tasks that require understanding both local features and their larger spatial context.

Finally, for more specialized applications, more complex, hybrid models may be necessary. For tasks where interpretability is a priority, incorporating latent-space constraints can improve both stability and clarity in the results. For example, using a Wasserstein GAN (WGAN) with a carefully designed loss function can provide more control over the training process and generate smoother, more interpretable transformations. In addition, hybrid models that combine multiple architectures, such as variational autoencoders (VAEs) with GANs, can provide both generative flexibility and the ability to impose structural constraints, improving the model’s capacity to generate accurate, interpretable results for complex tasks. In superresolution tasks, specialized models such as Deep-STORM or DECODE utilize domain knowledge to far outperform what standard cGANs can achieve.

Table 1

Overview of the key parameters for common approaches of DL in microscopy across scales.

6.3.

Evaluation Metrics

Evaluating the performance of a cross-modality transformation model can be challenging. Typical strategies involve measuring the visual fidelity of the images, but these measures may not fully correlate with the retention of biologically relevant information. Some examples of evaluation metrics include the following.

Another approach is to evaluate the biological relevance of the generated images by performing downstream analyses, such as cell counting or feature segmentation, and comparing the results to known quantities or those obtained from real experimental data. This method more directly assesses the retention of biologically meaningful information, but it introduces additional uncertainties. For example, inaccuracies in the downstream task, such as errors in the cell-counting algorithm, can confound the evaluation of the model’s performance, making it difficult to disentangle the model’s contributions from errors in postprocessing or analysis pipelines.

6.4.

Ethical Considerations

Ethical considerations are essential for ensuring responsible and fair use of AI in the image analysis of biological samples. A key concern is protecting the privacy of patients and donors, as these samples often contain sensitive personal information. Handling biological samples must comply with data protection laws, which require informed consent from all parties involved and transparency about how the samples will be used. These laws vary by region, with the General Data Protection Regulation (GDPR) in the European Union,¹⁷² the Health Insurance Portability and Accountability Act (HIPAA) in the United States,¹⁷³ the Data Protection Act 2018 in the United Kingdom,¹⁷⁴ and the Personal Information Protection Law (PIPL) in China.¹⁷⁵ International efforts, such as those led by the World Health Organization,^176,177 along with national initiatives, such as India’s data protection frameworks,¹⁷⁸ continue to evolve these regulations to keep pace with the rapid growth of AI technologies.^179–181 In addition, a growing body of literature addresses these developments across various countries,¹⁸² stages of database handling,¹⁸³ and specific fields.¹⁸⁴

These regulations address issues such as privacy breaches due to improper data use and cybersecurity threats.¹⁸⁵ Furthermore, the responsibility for AI models used in diagnostics is a major concern, particularly in relation to bias mitigation. Biased data sets can result in inaccurate or discriminatory outcomes, especially in healthcare applications. Rigorous validation of AI models is critical to ensure accuracy, reproducibility, and the prevention of errors that may lead to misdiagnosis or flawed scientific conclusions. Transparency is also crucial, requiring clear documentation of model training, data sources, usage frameworks, and decision-making processes.¹⁸⁶ Guidelines should promote not only sharing data sets but also the trained model weights, enabling researchers to independently validate and replicate findings. Lastly, accountability frameworks are necessary to ensure that researchers and developers are held responsible for the ethical use of AI, with proper oversight to enforce compliance with established guidelines.