Bridging scales: a perspective on macro-to-micro transfer for microscopy foundation models

Abstract

Foundation models (FMs) are transforming artificial intelligence for science, but microscopy FMs remain constrained by limited, heterogeneous, and information-sparse data. We propose a Macro-to-Micro paradigm in which large-scale macroscopic pretraining establishes a broad visual representation, followed by microscopy-specific adaptation that introduces biological specificity and modality-dependent imaging knowledge. This staged strategy enables scarce microscopy data to specialize a mature representation rather than build one from scratch. Combined with domain-aware adaptation and a coordinated microscopy data ecosystem, Macro-to-Micro provides a data-efficient roadmap toward foundation models that ground general visual intelligence in biological knowledge and the physics of image formation.

Modern microscopy reveals biological organization across scales, but the resulting images are shaped by photon budgets, diffraction, aberrations, staining protocols, detector noise, and modality-specific reconstruction¹. These constraints create trade-offs among resolution, speed, field of view, and signal-to-noise ratio². Their effects on training data differ across tasks and experimental settings. Some applications have extensive image collections but limited annotations; others lack representative images of particular structures, specimens, or acquisition conditions. These situations call for different learning strategies. In particular, limited annotation does not by itself imply a need for natural-image pretraining, because abundant unlabeled microscopy images may support in-domain self-supervised learning. The relevant question is whether the available data adequately represent the intended task and deployment conditions.

Here, as shown in Fig. 1, we propose Macro-to-Micro as a strategic paradigm for developing generalizable microscopy foundation models. Although macroscopic natural images and microscopic images differ in their physical origins and biological semantics, they share a rich repertoire of hierarchical visual structures, ranging from edges and textures to shapes and spatial organization³. Large-scale macroscopic pretraining can therefore establish a broad representational substrate that is difficult to learn from scarce and fragmented microscopy data alone, while microscopy-specific adaptation introduces biological specificity and reconciles modality-dependent imaging characteristics. Evidence from biomedical transfer learning and Segment Anything for Microscopy supports the feasibility of reusing and adapting general visual representations across domains^4,5. The key premise is thus not semantic equivalence between the macro- and microscopic worlds, but their representational complementarity. By combining the scale and diversity of macroscopic data with the specificity of microscopic data, the Macro-to-Micro paradigm provides a principled and data-efficient route toward more generalizable microscopy foundation models.

Fig. 1. Manifold alignment and the metaphorical strategy for data-efficient pre-training.

a Qualitative visualization of macroscopic natural images (blue) and microscopic images (orange) after feature extraction with a pretrained ResNet50 and projection using t-SNE. The apparent overlap is projection-dependent and is presented to motivate investigation of cross-domain representation transfer; it does not establish semantic equivalence or downstream transferability. b–d A metaphorical conceptualization of the FM training strategy. The cup symbolizes the capacity of the Foundation Model, while the stones represent universal structural priors from macroscopic data. Microscopic data is represented as water that infiltrates the interstitial gaps to provide functional specificity. The thirsty bird represents downstream biological tasks that become empowered through this fluid knowledge transfer, illustrating how the integration of both modalities overcomes data scarcity to achieve emergent generalization.

Pretraining options for microscopy foundation models

Several pretraining routes provide relevant points of comparison. Microscopy-domain pretraining learns directly from the target type of imagery, as illustrated by UniFMIR for fluorescence restoration⁶. Sequential natural-image and biomedical-image pretraining has also been used in RETFound, although its retinal imaging results should be distinguished from evidence in microscopy⁴. Synthetic-data approaches such as RETFound-DE and MINIM demonstrate uses of generated images for medical model training; their findings motivate, but do not establish, equivalent benefits in fluorescence or electron microscopy. Text can provide additional information. FluoResFM conditions fluorescence restoration on descriptions of tasks, structures, and imaging conditions⁷, while MUSK combines pathology images and text for oncology applications⁸. These strategies can be combined with different choices of visual pretraining.

The practical case for Macro-to-Micro is strongest when suitable natural-image encoders are available and comparable microscopy resources do not adequately cover the target setting. Reusing such encoders makes cross-domain transfer accessible without first constructing a microscopy-specific image generator. This convenience carries source-domain bias: natural-image features may emphasize patterns that are irrelevant or misleading in microscopy. Microscopy-only pretraining avoids that particular mismatch, although it remains sensitive to the coverage and biases of its own training data. Synthetic microscopy data offer control over selected structures and acquisition conditions, provided that the generation process preserves the features needed for the task. These differences support task-dependent comparisons rather than a general ranking of pretraining sources.

Several complementary strategies are already emerging. Microscopy- or biomedical-domain pretraining, as illustrated by fluorescence restoration and retinal foundation models^4–6, maximizes domain alignment but remains dependent on the scale and diversity of available in-domain data. Adaptation of generalist vision models can exploit broader visual diversity, but its success depends on how much of the pretrained representation is reusable and how the model is recalibrated^9,10. Synthetic-data pretraining can expand coverage of rare conditions and acquisition artifacts, although generator bias and imperfect physical realism may be inherited by the downstream model^11,12. Multimodal and vision-language models can connect images to semantic or clinical context⁷, but require reliable image-text alignment, consistent ontologies, and careful separation between pathology or medical imaging evidence and microscopy-specific evidence⁸. Macro-to-Micro transfer should therefore be understood as one component of a broader strategy, not as a replacement for microscopy-only pretraining, synthetic data, multimodal learning, or continued investment in microscopy datasets.

Why macro-to-micro: from sparse data to rich representations

Macro-to-Micro addresses a structural imbalance at the heart of microscopy foundation-model development. Microscopy datasets are limited not only in scale and diversity, but also in effective feature density: biologically decisive structures often occupy only a small fraction of an image, whereas background signals, preparation artifacts, and instrument-specific patterns account for much of the visual content¹³. Training exclusively on such data forces a model to learn both a general visual grammar and highly specialized biological semantics from the same scarce and heterogeneous samples. Large-scale macroscopic pretraining^14,15 separates these two learning burdens by first establishing a diverse and hierarchically organized representational foundation, which can then be refined and reweighted through microscopy-specific adaptation. Within this framework, the emergence of generalization depends on three coupled dimensions: the scale and diversity of the training data, the density of biologically informative features, and the efficiency with which learning objectives and model architectures transform those features into reusable representations. Macroscopic pretraining primarily expands representational breadth, whereas microscopic adaptation concentrates biological specificity and aligns the learned priors with modality-dependent imaging characteristics. Macro-to-Micro is therefore more than a change in initialization: it redistributes the learning burden across stages, enabling scarce microscopy data to specialize an already mature visual representation rather than construct one from the ground up.

A staged pipeline for macro-to-micro generalization

We formulate Macro-to-Micro learning as a progressive three-stage pipeline for constructing generalizable microscopy foundation models (Fig. 2). In the first stage, self-supervised pretraining on large and diverse macroscopic datasets enables a foundational encoder to internalize hierarchical visual regularities through objectives such as masked-image modeling¹⁴ or self-distillation¹⁵. This stage establishes a high-capacity representational substrate that would be difficult to construct from scarce microscopy data alone. In the second stage, limited but diverse microscopy data are used to build a domain-specific processor coupled to the pretrained encoder. Acting as the bridge between general visual priors and microscopic image formation, this processor introduces sensitivity to modality-dependent noise, scale, morphology, and biological context without requiring the model to relearn the entire visual vocabulary from the ground up. The macroscopic encoder can be retained to preserve representational breadth, while partial unfreezing, lightweight adapters, low-rank updates, or domain-aware normalization can provide additional flexibility when the source-target gap is large. In the third stage, the encoder–processor backbone is equipped with task-specific heads or prompts for restoration, segmentation, detection, tracking, and biological inference. This functional hierarchy is central to the paradigm: macroscopic pretraining builds representational capacity, microscopic adaptation directs biological specialization, and task-level supervision converts the resulting cross-domain representation into scientific utility.

Fig. 2. Architectural blueprint of the Macro-to-Micro training pipeline.

The framework organizes three functional stages designed to address data scarcity in microscopy. First, large-scale macroscopic data support self-supervised pretraining of the foundational encoder. Second, limited but diverse microscopy data are used for domain-specific self-supervised adaptation through a microscopy processor, with the encoder retained or selectively updated according to domain distance. The resulting encoder–processor backbone is then adapted to downstream tasks using task-specific heads or prompts.

This staged architecture also generates concrete and falsifiable predictions. First, the advantage of macroscopic pretraining should become more pronounced as the amount and diversity of available microscopy data decrease. Second, transfer performance should depend not only on the source encoder, but also on the ability of the microscopy processor to reconcile general visual structure with modality-specific imaging physics. Third, the optimal balance between preserved macroscopic priors and adaptive microscopic capacity should vary systematically with domain distance. These predictions can be examined by comparing Macro-to-Micro and microscopy-only pretraining under matched data and computational budgets, evaluating transfer across imaging modalities and downstream tasks, and quantifying the preservation of rare biological structures. Complementary analyses should distinguish genuine feature reuse from optimization or weight-scaling effects and assess possible texture bias and acquisition-specific shortcuts. Figure 2 therefore represents more than a collection of implementation choices: it provides an architectural blueprint and an empirical roadmap for transforming cross-domain transfer into a generalizable microscopy foundation-model strategy.

Domain mismatch and negative transfer

Macro-to-Micro is designed to operate across heterogeneous microscopy modalities without assuming that every modality requires the same depth of adaptation. Its strength lies in explicitly separating representations that can be transferred from those that must be reconstructed within the microscopic domain. Hierarchical visual priors related to edges, textures, shapes, and spatial organization may remain broadly reusable, whereas modality-dependent noise statistics, image-formation physics, and biological semantics require increasingly specialized adaptation. Negative transfer can arise when natural-image texture priors overshadow biologically meaningful morphology, when rare structures have few macroscopic analogues, or when the measured signal is governed more strongly by imaging physics than by generic visual appearance. These challenges are particularly relevant to super-resolution reconstruction, cryo-electron microscopy, phase imaging, and other physics-sensitive modalities. Rather than constituting exceptions to the paradigm, such cases define regimes in which the microscopy-specific processor must assume a greater representational burden and the macroscopic encoder must be adapted more selectively.

Source selection and adaptation depth should be chosen using target-domain validation. Comparisons can test frozen encoders, partial or full fine-tuning, adapters, and domain-specific normalization under the same downstream evaluation conditions. Their ability to reduce negative transfer should be assessed against microscopy-only and, where appropriate, synthetic-data pretraining. Evaluation should include rare structures and held-out instruments or acquisition conditions, together with checks for reliance on staining, texture, or laboratory-specific patterns. Restoration and reconstruction tasks also require tests of whether biologically relevant signals are preserved and spurious structures are introduced. Aggregate performance alone may conceal failures on uncommon targets. A source representation should be retained when these comparisons support its use; evidence of persistent negative transfer should guide the choice toward a different pretraining source or learning strategy.

Scaling macro-to-micro: toward a global microscopy data ecosystem

Realizing the full potential of Macro-to-Micro will require a microscopy data infrastructure capable of supporting systematic adaptation across biological scales, imaging modalities, instruments, and laboratories. Current repositories remain fragmented in acquisition protocols, metadata quality, file formats, annotation standards, and biological coverage, limiting the ability of models to distinguish conserved biological structure from instrument- or protocol-specific artifacts. A global microscopy data ecosystem should therefore integrate standardized acquisition metadata, interoperable formats, calibrated physical parameters, and ontology-driven annotations that connect morphology with molecular, functional, and experimental context. Such an ecosystem would not merely increase the volume of microscopy data; it would organize microscopic diversity into a representation space that can be effectively aligned with, and differentiated from, knowledge learned from the macroscopic world.

Scaling Macro-to-Micro depends on expanding the microscopy data available for adaptation and validation. Macroscopic pretraining offers a pragmatic starting point when microscopy data cannot yet support pretraining at the required scale and diversity. Microscopy data remain necessary to learn biological specificity and modality-dependent imaging characteristics, and to establish whether transferred representations improve downstream performance. A unified microscopy data ecosystem would support this process through consistently documented datasets spanning instruments, laboratories, and experimental conditions. Such coverage would enable rigorous, multidimensional evaluation of task performance, biological fidelity, and generalization, including failures on rare structures and unfamiliar acquisition conditions. Real observations from extreme imaging regimes and emerging modalities would also guide the development of augmentation and synthesis methods, with independent measurements used to assess their physical and biological validity. These resources could extend the range of microscopy conditions represented during adaptation and reveal where macroscopic priors become inadequate. As microscopy datasets grow sufficiently large and representative, the additional value of macroscopic pretraining may diminish. Building this ecosystem therefore supports the reliable use of Macro-to-Micro under current data constraints while developing the resources needed for more extensive learning directly from microscopy data.

Author contributions

Y.G. and Y.W. performed the literature review and data curation. B.D., J.Z. and Z.L. supervised the project, provided critical expertise, and revised the manuscript. All authors read and approved the final version.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

Biqin Dong is a founder and equity holder of MicroLux (Shanghai) Intelligent Science & Technology Co., Ltd. and Feimiao Intelligent Science & Technology (Taizhou) Co., Ltd. All other authors declare no competing interests.