Computational Miniaturized Mesoscope (CM²)

• CM² is a compact fluorescence imaging platform that integrates a 3×3 microlens array with advanced computational reconstruction for high-resolution, volumetric imaging.
• It employs a unique optical design and model-based plus deep-learning algorithms to achieve micron-scale lateral resolution over centimeter-scale fields of view.
• The system addresses traditional limitations in space-bandwidth product, depth-of-field, and aberrations, enabling diverse applications from rodent brain imaging to on-chip cytology.

The Computational Miniaturized Mesoscope (CM²) is a compact, single-shot, computational fluorescence microscopy platform that leverages a microlens array (MLA) to spatially multiplex multiple perspective views of an object onto a single image sensor. By combining this optical architecture with advanced model-based and deep-learning computational reconstruction pipelines, CM² achieves micron-scale lateral resolution and volumetric (3D) imaging over centimeter-scale fields of view, previously unattainable in such a miniaturized form factor. The technology addresses fundamental limitations of conventional miniaturized microscopes—particularly space-bandwidth product (SBP), depth-of-field (DOF), and spatially varying aberrations—by tightly co-designing optics and algorithms for high-throughput, robust fluorescence imaging in systems constrained by size, weight, and power (Xue et al., 2020, Xue et al., 2022, Yang et al., 2024, Yang et al., 1 Feb 2026).

1. Optical Architecture and Physical Principles

At the core of the CM² platform is a 3×3 square MLA (pitch ≈ 1 mm; focal length 1.2–7.5 mm, depending on variant) positioned at a finite conjugate above a single planar, back-illuminated CMOS sensor. Each of the nine microlenses forms a distinct, low-NA image of the object, encoding depth as a characteristic lateral shear between the perspectives. Fluorescence excitation is provided by surface-mount LEDs (470/40 nm bandpass) with custom collimators and hybrid interference/absorption emission filters, yielding up to 5× contrast and 3× excitation efficiency improvements in subsequent versions (CM² V2) (Xue et al., 2022).

All optical components—including the MLA, filters, LEDs, and the sensor—are integrated in a lightweight, custom 3D-printed enclosure. There are no conventional objectives, dichroic cubes, or relay optics. This integration eliminates bulky infinite-conjugate elements and GRIN rods found in miniscopes, enabling total instrument footprints as small as 20×20×13 mm³ and weights under 2.5 g (Xue et al., 2022). Field of view is typically 6.5–8 mm, and depth-of-field extends up to 2.5 mm, with lateral resolution improved by computational processing to the 6–8 μm regime.

2. Computational Forward Models and Depth Encoding

Fluorescence imaging measurements in CM² are mathematically described by spatially varying convolution with depth-dependent point spread functions (PSFs):

$$I(x, y) = \iiint H(x, y, z; x', y', z')\, O(x', y', z')\, dx'\, dy'\, dz' + \eta(x, y).$$

By discretizing into axial slices and imposing shift-invariant approximations, single-shot 3D images are modeled as the sum over depth-slice convolutions:

$$I = \sum_{k=1}^{N_z} (\mathrm{PSF}_k * O_k) + \eta,$$

with $O_k$ denoting object slices and $\mathrm{PSF}_k$ their corresponding system PSFs (Xue et al., 2020).

Physically, the 3×3 MLA imparts a distinct axial shear to the image array with object depth, so $d(z) = \frac{D}{l_0} \Delta z$ (with $D$ the MLA pitch and $l_0$ the MLA–sensor distance). This axial shearing rapidly decorrelates the multiview PSFs as depth varies (Pearson correlation <0.8 beyond ≈155 μm), providing depth localization without mechanical scanning. In advanced implementations, these PSFs are measured densely on a grid and compressed via truncated SVD or other low-rank decompositions, forming the basis of fast forward simulation and inverse reconstruction (Xue et al., 2022, Yang et al., 2024).

3. Inverse Algorithms: Model-Based and Learning-Based Reconstruction

Model-Based Deconvolution

Initial CM² protocols adopted a regularized least-squares deconvolution framework with both sparsity-promoting $\ell_1$ and 3D total variation regularizers, subject to a non-negativity constraint:

$$\hat x = \arg\min_{x \geq 0} \; \frac{1}{2}\| y - D H x \|_2^2 + \lambda_1\|x\|_1 + \lambda_{\text{TV}} \mathrm{TV}_{3D}(x)$$

where $D$ is a truncation operator and $H$ encodes the 3D convolution. ADMM-based optimization, using proximal and FFT-accelerated substeps, yields reconstructions with lateral resolution $\sim$7–11 μm and axial localization $<$200 μm, though with computational cost of hours per 3D volume on CPUs (Xue et al., 2020).

Deep Learning Augmentation

Recent advances supplant iterative deconvolution with dedicated deep neural networks. The CM²Net (CM² V2) pipeline is a three-branch network:

1. View-Demixing Net (12-layer residual block stack): learns to unmixed crosstalked multi-view images based on lens-specific aberration fingerprints.
2. View-Synthesis Net (8 residual blocks): directly infers 3D volume from parallax across the nine views via upsampling in $z$.
3. Refocusing-Enhancement Net (5-stage U-Net): sharpens lateral detail and mitigates axial elongation via geometric refocusing.

Training leverages only synthetic data, with loss as the sum of binary cross-entropy terms for demixing and final volumetric prediction. CM²Net achieves $\sim$6 μm lateral, $\sim$25 μm axial accuracy across $\sim$7 mm FOV, with $<$4 s inference per 230 M-voxel 3D volume on commodity GPUs (Xue et al., 2022).

State-of-the-Art Spatially Varying (SV) Deep Architectures

To address field-dependent aberrations and scalability, two recent deep learning methods stand out:

• SV-FourierNet (Yang et al., 2024): A two-stage Fourier-domain neural network with learnable complex-valued deconvolution filters and a lightweight residual channel attention network. Suitable for 2D imaging, SV-FourierNet efficiently handles severely non-local and spatially varying PSFs, achieving 7.8 μm resolution and real-time, full-FOV inference (0.1 s per frame for 12 MP images).
• SV-CoDe (Spatially Varying Coordinate-conditioned Deconvolution) (Yang et al., 1 Feb 2026): Employs coordinate-conditioned convolutions where compact MLPs generate per-channel gating masks as a function of spatial coordinates, enabling efficient per-patch (480×480 px) scalable processing. This modularity decouples parameter count and memory from total FOV, facilitating training and inference for gigapixel-scale data. SV-CoDe surpasses previous baselines (including SV-FourierNet) in PSNR and SSIM while requiring 10x fewer parameters and training samples.

4. Performance Characterization and Quantitative Metrics

Comprehensive experimental and simulated validation demonstrates:

• Resolution: CM² systems consistently achieve lateral FWHM between 6–8 μm (USA target, bead phantoms) and axial FWHM as low as 25 μm (CM²Net), depending on algorithm and optical configuration.
• Field of View/Depth of Field: Uniform performance over FOVs of 6.5–8 mm diameter and DOF up to 2.5 mm, with minimal edge degradation (recall of $\sim$0.35 at 3.5 mm radius; central precision/recall $\sim$0.7–0.9) (Xue et al., 2022, Yang et al., 2024).
• Processing Speed: Model-based deconvolution requires hours/volume; CM²Net and SV-FourierNet reduce this to seconds or below per frame, with SV-CoDe (2.3M parameters) maintaining rapid scalability (Yang et al., 1 Feb 2026).
• Experimental Generalization: Trained entirely with physics-based simulators, modern deep-learning pipelines generalize robustly to 10 μm bead phantoms, weakly scattering brain slices, and dynamic in vivo imaging of C. elegans colonies without retraining (Yang et al., 1 Feb 2026, Yang et al., 2024).

System/Algorithm	Lateral Resolution	FOV	Speed (frame, GPU)	Model Size
Model-based (ADMM)	7–11 μm	8×7 mm²	2.5 h (3D volume)	—
CM²Net	6 μm / 25 μm (ax.)	7 mm diam	3.6 s (volumetric)	12M params
SV-FourierNet	7.8 μm	6.5 mm diam	0.1 s (2D frame)	1.8M params
SV-CoDe	7.8 μm	6.5 mm diam	0.59 s (2D frame)	2.3M params

5. Biological and Biomedical Applications

CM² enables wide-area, cellular-resolution fluorescence imaging in domains requiring minimal device mass and footprint. Demonstrated and prospective applications include:

• Cortex-wide 3D calcium imaging in freely behaving rodents, exploiting the 7 mm FOV to cover mouse cortex (Xue et al., 2020, Xue et al., 2022).
• Whole-brain imaging in larval zebrafish, where near-micron resolution and single-shot 3D recovery support large-volume studies.
• High-throughput functional imaging of C. elegans colonies, with dynamic video-rate reconstructions faithfully tracking motile worms (Yang et al., 2024, Yang et al., 1 Feb 2026).
• Histological studies (e.g., mouse brain coronal sections), where resolved 2–5 μm cellular structures compare favorably with benchtop widefield references.
• On-chip cytology, lab-on-a-chip diagnostics, and endoscopic imaging, leveraging the compact, lensless geometry (Yang et al., 2024).

6. Trade-Offs, Limitations, and Future Directions

Key trade-offs in CM² include decreased photon efficiency ($\sim$20%) and multiplexing-induced crosstalk as spatial information is distributed across multiple sub-views. While model simplifications (e.g., slice-wise shift-invariance) yield faster computation, they can incur minor resolution losses near FOV boundaries. Higher performance reconstruction networks such as SV-CoDe mitigate these effects with spatially adaptive processing but at increased computational complexity (Yang et al., 1 Feb 2026).

Current limitations are posed by the use of one-photon excitation (susceptible to scattering), side-view occlusion at extreme field positions, and restriction of current neural networks to sparse and bead-like structures. Future directions target volumetric spatially varying deconvolution (extending coordinate conditioning to $(x, y, z)$), deep learning approaches for continuous structure imaging, integration of structured illumination, scattering-aware forward models, and further miniaturization using custom sensors/LEDs for in vivo head-mounted deployment (Xue et al., 2022).

A plausible implication is that coordinate-conditioned or global-frequency-domain networks offer a generalizable framework for correcting spatially varying aberrations in compact instruments outside the CM² class—such as light-field microscopes, metalens-based endoscopes, and camera arrays—provided accurate physical models are incorporated (Yang et al., 1 Feb 2026).

7. Comparative Analysis and Significance

The CM² architecture fundamentally alters the FOV–resolution–complexity balance in miniature microscopy. By shifting traditional optical constraints into the computational domain and co-designing specialized deep algorithms, it achieves uniform, near-diffraction-limited performance in a highly constrained form factor. End-to-end learning entirely on simulated physics-based data—without empirical PSF libraries—yields networks that robustly generalize across diverse experimental modalities.

Relative to prior art, CM² and its associated networks:

• Provide $\sim$8× axial resolution improvement and $\sim$10–1000× speedup over model-based deconvolution (Xue et al., 2022).
• Demonstrate consistent performance across entire wide fields, outperforming low-rank and classical spatially varying deconvolution baselines in PSNR and SSIM (Yang et al., 1 Feb 2026, Yang et al., 2024).
• Require an order of magnitude fewer trainable parameters and training samples than conventional spatially varying architectures, contributing to practical scalability on modern hardware.

The resulting computational miniaturized mesoscope platform represents a significant advance in optics–computation co-design for high-throughput, large-area biological imaging (Xue et al., 2020, Xue et al., 2022, Yang et al., 2024, Yang et al., 1 Feb 2026).