HAMscope: Dual-Mode High-Contrast Imaging

• HAMscope is a dual-mode imaging system featuring holographic aperture masking for astronomy and a compact hyperspectral miniscope for bioimaging.
• The astronomical design uses liquid-crystal geometric phase masks to multiplex subapertures, improving throughput and phase calibration in high-contrast interferometry.
• The bioimaging system employs a probabilistic multi-pass U-Net for rapid spectral unmixing, enabling real-time quantitative mapping of molecular emissions.

HAMscope refers to two distinct high-performance optical and computational systems for interferometric and hyperspectral imaging: (1) a multiplexed holographic aperture masking architecture leveraging liquid-crystal geometric phase masks for high-contrast astronomical imaging (Doelman et al., 2018), and (2) a compact snapshot hyperspectral autofluorescence miniscope for real-time molecular imaging in biological contexts (Ingold et al., 12 Nov 2025). Both utilize hardware minimization coupled with advanced computational algorithms to deliver quantitative imaging at high spatial and spectral fidelity, but target fundamentally different application domains.

1. Physical Principles and Optical Architectures

1.1. Holographic Aperture Masking (“HAMscope” in astronomy)

HAMscope implements holographic multiplexing of sparse pupil subapertures via spatially patterned liquid-crystal half-wave retarders generating spatially controlled Pancharatnam–Berry geometric phase shifts. For right-/left-circular polarization (σ=±1), the emergent phase is φ_g(x,y) = σ·2θ(x,y), where θ(x,y) is the local fast-axis orientation. The patterned liquid-crystal layer is written at micron pitch, with net retardance tuned to half-wave at the operation band to maximize diffraction efficiency and phase uniformity. Light from each selected subaperture is redirected into multiple, non-redundant off-axis PSFs through a sequence of blazed holographic gratings. The focal plane intensity for the k-th order is

$$I_k(x', y') = | \mathcal{F}\{ A(x, y) \exp{i[\varphi(x, y) + c_k]} \} \otimes \delta(x' - x'_k, y' - y'_k) |^2,$$

where A(x, y) selects the subaperture, c_k is a piston bias, and (x'_k, y'_k) sets the PSF position. Arbitrary subaperture combinations and PSF configurations are programmable via the holographic phase profile.

1.2. Snapshot Hyperspectral Miniscope (“HAMscope” in bioimaging)

HAMscope, for label-free molecular imaging, retrofits a widefield Miniscope V4 with a thin polymer diffuser (scatter angle ≈0.046°) at the image plane. Under UV-LED excitation, the diffuser imparts a wavelength-dependent, spatially varying point spread function (PSF) to every image. A reference 4f relay allows simultaneous acquisition of hyperspectral ground truth (via a linear-variable bandpass filter, LVBF) and encoded monochrome images. The forward imaging model is

$$\mathbf{y} = \sum_{k=1}^{N_\lambda} H^{(\lambda_k)} \mathbf{x}^{(\lambda_k)} + \boldsymbol{\eta}$$

where $\mathbf{x}^{(\lambda_k)}$ is the object’s emission at wavelength $\lambda_k$, $H^{(\lambda_k)}$ is the measured PSF stack, and $\boldsymbol{\eta}$ is sensor noise. Calibration of the full PSF set yields the system response across all channels.

2. Computational Methods for Inverse Imaging

2.1. Probabilistic Deep Networks for Spectral Reconstruction

In the hyperspectral HAMscope, the spectral unmixing problem $y \rightarrow \hat{x}$ is solved using a probabilistic, multi-pass U-Net architecture with transformer-based self-attention at the bottleneck. Multiple passes (P=3) recursively deepen features via shared skip connections, while an 8-head transformer improves spatial–spectral context. The output at each pixel is a Laplacian distribution parameterized by mean $\mu_{i,k}$ and scale $\sigma_{i,k}$ over spectral bands:

$$\{ \boldsymbol{\mu}, \boldsymbol{\sigma} \} = G_\theta(\mathbf{y})\,,\ \ \ p(x_{i,k}|y) = \frac{1}{2 \sigma_{i,k}} \exp\left( -\frac{|x_{i,k} - \mu_{i,k}|}{\sigma_{i,k}} \right).$$

The loss combines negative log-likelihood, $L_1$ error, and a weak adversarial (least-squares) discriminator:

$$\mathcal{L}_{\rm total} = \mathcal{L}_{\rm NLL} + \lambda_1 \|\boldsymbol{\mu} - \mathbf{x}\|_1 + \lambda_{\rm adv}\, \mathbb{E}_{\mathbf{y}} \left[ \| D_\phi(G_\theta(\mathbf{y})) - 1 \|_2^2 \right]$$

Per-pixel uncertainty estimates (aleatoric and epistemic) are derived from $\sigma_{i,k}$ and ensemble KL divergence, respectively.

2.2. Holographic Mask Design and PSF Synthesis

The holographic HAMscope achieves tailored uv-coverage by optimizing subaperture selections across multiple PSF copies. Each mask multiplexes $K$ non-redundant baselines by summing $K$ blazed-hologram phase patterns. This is captured formally as

$$\phi_h(x, y) = \mathrm{arg}\left[ \sum_{k=1}^K s_k \exp\big(i [2\pi(f_{kx} x + f_{ky} y) - c_k] \big) \right]$$

where $s_k$ are order weights, $f_{kx},f_{ky}$ are spatial frequencies, and $c_k$ is a piston. The union of all baseline vectors $b_{ij} = r_i - r_j$ (for subaperture centers $r_i$) across multiplexed masks augments uv-coverage.

3. Quantitative Performance and Metrics

3.1. Hyperspectral Miniscope (Biological HAMscope)

• Spatial Resolution: 10.22 µm in raw reconstructions at 542 nm, improved to 6.74 µm post-deconvolution (Fourier cutoff: 0.055 µm⁻¹, ground truth 9.13 µm).
• Spectral Coverage: 30 bands from 452 to 703 nm (Δλ ≈ 8.5 nm).
• Speed: 10 Hz hyperspectral imaging (single-frame decoding in 0.1 s, reference system ≈141 s/frame).
• Accuracy: Triple-pass probabilistic U-Net achieves mean absolute error (MAE) ≈ 0.0048 per pixel/channel.
• Out-of-distribution Generalization: Maintains MAE ≈ 0.0718 on cork oak (suberin), with network uncertainty flagging missing features in 650–700 nm band.

3.2. Holographic Aperture Masking (Astronomical HAMscope)

• Throughput: Transmission $$T_{\mathrm{HAM}} = K\, n_{\text{sh}} (D_s/D_{\text{total}})^2$$, a factor $K$ larger than conventional SAM ($T_{\mathrm{SAM}}$), for $K$ non-redundant PSF copies.
• Raw Contrast: 1 × 10⁻³ at 1 λ/D; 5 × 10⁻⁴ at 3 λ/D (measured at 532 nm).
• Closure-phase Scatter: RMS ≈ 0.2° (improves to 0.15° with amplitude-calibrated reference spots).
• Fringe SNR: Exceeds 100 per baseline in 1 s integrations.

A summary of key instrument metrics is provided below:

Parameter	Astronomical HAMscope (Doelman et al., 2018)	Bioimaging HAMscope (Ingold et al., 12 Nov 2025)
Spatial Resolution	λ/Dₛ (astronomy, e.g. 0.7″ at 532 nm)	10.22 μm raw, 6.74 μm deconvolved at 542 nm
Spectral Channels	Typically broadband	30 (452–703 nm, Δλ ≈ 8.5 nm)
Throughput	Up to 4× conventional masks (K=4)	Full FOV per snapshot
Application Domain	High-contrast small-angle interferometry	Real-time, label-free autofluorescence

4. Application Domains and Demonstrated Use Cases

4.1. Biological Imaging

HAMscope provides label-free, real-time mapping of endogenous biomolecules in plant tissues, including lignin (∼475 nm emission), chlorophyll (650–750 nm), and suberin. Applications demonstrated include:

• Temporal monitoring of branch wound healing (poplar): 10 Hz imaging over 43 hours tracking lignin and chlorophyll dynamics.
• Air–water interface motion in xylem vessels (embolism input) of transgenic poplar.
• Cork oak bark: direct suberin detection, spectrally separable from lignin/chlorophyll.

A plausible implication is that this architecture, with minimal hardware change and computational retraining, is transferable to neural imaging (e.g. spectral indicators), metabolic mapping (NADH/FAD), and histopathology (margin delineation with endogenous fluorophores).

4.2. High-Contrast Astronomical Imaging

In astronomical HAMscope, programmable multiplexed phase masks yield enhanced throughput and uv-plane coverage for non-redundant masking interferometry. The apparatus is compatible with focal-plane wheels and relays of telescopes scaling up to ~100 mm pupil size. On-sky deployment enables order-of-magnitude higher photon yields than classical SAM, supporting high-SNR, self-calibrating closure-phase interferometry inside 1 λ/D.

5. Direct End-to-End Molecular Mapping

HAMscope’s learning pipeline supports direct regression from monochrome, diffuser-encoded images to per-pixel molecular abundance maps:

$$\hat{\mathbf{m}} = G_\psi(\mathbf{y}) \qquad \mathbf{m}_{i,c} = \sum_{k=1}^{30} w_{c,k}\, x_{i,k}$$

where $w_{c,k}$ are reference spectra for each molecular component. End-to-end training via L1 loss

$$\mathcal{L}_{\rm map} = \| \hat{\mathbf{m}} - \mathbf{m} \|_1$$

yields MAE ≈ 0.0064 (single step; improved to ≈0.0010 with probabilistic U-Net), outperforming two-step “hyperspectral→spectral unmixing” pipelines. This supports direct, quantitative biochemical mapping across complex samples in field or laboratory settings.

6. Scalability and Practical Implementation

• Hardware Integration: Astronomical HAMscope is a single transmissive optic (few mm thick), readily scalable to large-aperture relays. Retardance uniformity better than ±1% demonstrated over 50 mm apertures.
• Computational Load: Hyperspectral HAMscope achieves 10 Hz hyperspectral reconstruction on mid-size GPUs (NVIDIA RTX 3060); U-Net parameters increase sub-linearly with multi-pass design.
• Detector Constraints: Astronomical architectures are limited by detector area (each PSF ~64 effective pixels per copy, K_max ∝ A_det/(8λ/Dₛ)²). For plant miniscope, full sensor area is used per snapshot.

A plausible implication is that future deployments can modularly exploit advances in camera technology, computational accelerators, and adaptive optics, further increasing field utility and imaging fidelity.

7. Summary

HAMscope designates both (1) a holographically multiplexed, geometric-phase-based aperture masking system for high-contrast, small-angle astronomical imaging (Doelman et al., 2018) and (2) a minimal-optics, probabilistic deep-learning-enabled snapshot hyperspectral miniscope for real-time label-free molecular imaging in biology (Ingold et al., 12 Nov 2025). Both systems achieve substantial gains in throughput and calibration or spatio-spectral fidelity by merging hardware-encoded multiplexing with advanced statistical reconstruction. Field and laboratory studies corroborate high-resolution, high-fidelity, uncertainty-aware performance in both domains, with broad applicability spanning interferometric astronomy to molecular bioimaging and potential for rapid translation across scientific disciplines.