🔬 optics

Numerical-aperture transfer in holotomography with a deterministic diffusion prior

This paper introduces ResShift-ODE, a deterministic diffusion-prior framework that computationally transfers low-numerical-aperture refractive-index holotomography data to high-NA-equivalent volumes with high accuracy and reproducibility, achieving results comparable to high-NA references in just five denoiser evaluations without requiring hardware modifications.

Original authors: Dong-min Ryu, Jeong-hyun Roh, Woon-soo Lee, Hansol Yoon, Hyun-seok Min, Yongkeun Park

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dong-min Ryu, Jeong-hyun Roh, Woon-soo Lee, Hansol Yoon, Hyun-seok Min, Yongkeun Park

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a high-definition, 3D photo of a tiny, living cell to see its internal machinery. In the world of microscopy, there is a trade-off: to get a super-sharp, detailed picture (high resolution), you usually need a lens that is very close to the sample. But if you want to look at many samples at once—like in a multi-well plate used for drug testing—you need a lens that sits further away. This "long-distance" lens is great for speed and volume, but it acts like a pair of foggy glasses: it blurs out the tiny details and loses the fine textures of the cell.

The paper introduces a new software tool called ResShift-ODE that acts like a "digital lens upgrade." It takes those blurry, low-detail 3D images and mathematically reconstructs them to look as sharp and detailed as if they had been taken with the expensive, close-up lens—without ever changing the physical microscope.

Here is how it works, using simple analogies:

1. The Problem: The "Missing Puzzle Pieces"

Think of a high-resolution image as a complete jigsaw puzzle. When you use the low-quality lens, it's as if someone took the puzzle, removed all the pieces from the center and edges, and only gave you the pieces from the middle.

The Old Way: Previous computer programs tried to guess the missing pieces by just "smoothing out" the blurry edges or trying to guess what the picture might look like. Sometimes they guessed wrong, inventing details that weren't there (hallucinations).
The New Way: This paper treats the missing pieces as a specific "band" of information that is completely gone, not just blurry. The goal isn't to fix a blur; it's to fill in the missing puzzle pieces based on a learned pattern, while strictly keeping the pieces you do have exactly as they are.

2. The Solution: A "Deterministic Diffusion" Recipe

The researchers used a type of AI called a Diffusion Model. You can think of a standard diffusion model like a game of "Telephone" played in reverse:

The Standard Game: You start with a clear picture and slowly add static noise until it's just white fuzz. Then, the AI tries to guess how to remove the noise step-by-step to get the picture back.
The ResShift Twist: Instead of starting with white fuzz, this new method starts with your blurry, low-quality image and "shifts" it toward the sharp, high-quality version. It's like having a blurry photo and asking the AI, "If we knew what the sharp version looked like, what is the most logical path to get there?"

3. The "ODE" Shortcut: The Fast Lane

Usually, these AI models take a long time to work because they have to take hundreds of tiny steps to remove the noise (like walking up a staircase one step at a time).

The Innovation: The authors reformulated the math into a "Probability-Flow ODE." Think of this as finding a high-speed elevator instead of walking up the stairs.
The Result: Instead of taking 1,000 steps (which takes hours), the new method takes just 5 steps (which takes a few minutes). Crucially, because it's a "deterministic" path (like a train on a fixed track), if you run it twice, you get the exact same result every time. This makes the results reliable and reproducible.

4. What It Actually Does (and Doesn't Do)

The paper tested this on three types of cells: yeast, liver cells (HepG2), and blood cells (K562).

The Success: The software successfully added back the fine textures, sharp boundaries, and internal dots (puncta) that were missing in the low-quality images. It made the blurry cells look like high-definition 3D models.
The Safety Check: The researchers were careful to ensure the AI didn't "make things up." They checked the math in the "frequency domain" (a way of looking at the image's data patterns). They found that the AI only filled in the missing side-to-side details. It did not try to invent information in the "missing cone" (a specific 3D shape of data that is physically impossible to capture with this type of microscope). It respected the physical limits of the hardware.

5. The Bottom Line

This paper presents a software upgrade that allows scientists to use fast, long-distance microscopes (great for screening many samples) and then use this AI to instantly upgrade the images to look like they came from a slow, high-magnification microscope.

Speed: It is about 166 times faster than previous high-end AI methods.
Accuracy: It recovers details with very high precision (99% of the image pixels are within a tiny error margin).
Reliability: It gives the same answer every time you run it.

The authors emphasize that this was tested on "emulated" data (where they mathematically simulated the blur on high-quality images to train the AI). While the results are promising, the paper notes that the next step is to prove it works perfectly on real-world, physically blurry images taken from actual microscopes. However, the method offers a powerful new way to get high-resolution 3D cell data without needing expensive hardware upgrades.

Technical Summary: Numerical-Aperture Transfer in Holotomography with a Deterministic Diffusion Prior

Problem Statement
High-throughput holotomography (HT) faces a fundamental trade-off between screening compatibility and subcellular resolution. While research-grade systems using short-working-distance objectives achieve high illumination and detection numerical apertures (NA), high-throughput applications (e.g., multiwell plates, organoid screens) require long-working-distance optics that limit illumination NA (often <0.5). This reduction truncates the accessible 3D Fourier support of the scattering potential, suppressing high-spatial-frequency refractive index (RI) content, broadening fine structures, and reducing apparent contrast.

Existing computational approaches, such as deep learning regularizers or projection-domain learning, address image quality or missing-cone artifacts but are not designed for NA transfer: recovering high-NA-equivalent 3D RI volumes from low-NA measurements under a known, NA-limited Fourier-domain forward operator. The core challenge is that the missing high-frequency band is not merely blurred but entirely absent from the measurement, making the inverse problem ill-posed. The task requires a learned prior to estimate unmeasured frequency content while remaining strictly anchored to the measured low-NA support to avoid hallucination.

Methodology: ResShift-ODE
The authors propose ResShift-ODE, a deterministic diffusion-prior framework for transferring low-NA RI tomograms to high-NA-equivalent volumes without hardware modification.

Physical Formulation: The problem is formulated as the conditional recovery of a high-NA volume ( $x_H$ ) from a low-NA observation ( $x_L$ ). The forward operator is defined as a coherent Fourier-support restriction: $x_L = \mathcal{F}^{-1}[M_{K_L} \mathcal{F}[x_H]] + n_{obs}$ , where $M_{K_L}$ is the mask for the low-NA synthetic aperture.
Data Generation: To create voxel-registered training pairs, the authors apply the analytic low-NA forward operator to experimentally acquired high-NA reconstructions. This "low-from-high" strategy emulates the full measurement-and-reconstruction chain, including non-linear regularization, to produce a realistic surrogate for physical low-NA data.
Diffusion Prior (ResShift): Unlike standard Denoising Diffusion Probabilistic Models (DDPMs) that diffuse toward an unconditional Gaussian, ResShift employs a residual-shifting Markov chain. The forward process shifts the high-NA state toward the low-NA measurement ( $x_L$ ) rather than noise. This anchors the reverse trajectory to the observed data.
Deterministic Sampling (ODE): To enable reproducible, few-step inference, the stochastic reverse process is reformulated as a probability-flow Ordinary Differential Equation (ODE). This removes the Brownian motion term while reusing the trained denoiser. The reverse trajectory is integrated using a first-order Euler scheme over five steps (5 denoiser evaluations).
Network Architecture: The model utilizes a customized 3D U-Net backbone featuring a factorized Conv2D + axial-attention core to manage 3D memory constraints, pre-activation residual blocks, and rotary position embeddings. The input concatenates the current diffusion state with the low-NA condition.

Key Results
The method was evaluated on held-out emulated-pair test volumes spanning three cell types: yeast (S. cerevisiae), adherent HepG2 cells, and K562 lymphoblasts.

Reconstruction Accuracy: On held-out volumes, the inferred high-NA-equivalent outputs matched high-NA references with RI errors of 0.002–0.003 for >99% of voxels. Specifically, 99.92% of yeast voxels, 99.39% of HepG2 voxels, and 99.51% of K562 voxels fell within these error bounds.
Fourier-Domain Behavior: Fourier analysis confirmed that the model successfully recovered spatial frequencies in the lateral high-frequency annulus (outside the low-NA support but within the high-NA support). Crucially, the model did not artificially fill the axial missing cone, which remains empty in both the low-NA input and the high-NA reference. This demonstrates that the prior completes only the lateral band populated in the high-NA targets while respecting the physical constraints of the missing cone.
Computational Efficiency: ResShift-ODE requires only 5 denoiser evaluations per volume.
- It is approximately 166× faster than a 1000-step 3D DDPM baseline.
- It is approximately 2.7× faster than stochastic 3D ResShift (which requires 15 steps).
- It incurs a computational cost of ~5.8× that of a deterministic 3D U-Net, but provides a generative prior for band completion rather than simple regression.

Significance and Claims
The paper claims that the primary contribution is not the diffusion sampler itself, but the problem formulation: treating low-NA-to-high-NA transfer as a diffusion-prior inverse problem under an explicit, NA-limited Fourier-domain forward operator.

Measurement-Anchored Completion: The framework achieves "band completion" rather than deconvolution. By anchoring the reverse trajectory to the low-NA measurement, the method ensures that recovered structures in the unmeasured high-frequency band are prior-guided estimates constrained by the data, rather than indiscriminate synthesis.
Reproducibility: Reformulating the process as a probability-flow ODE enables deterministic, reproducible inference in just five steps, addressing the stochasticity and high computational cost of standard diffusion models.
Software-Only Resolution Enhancement: The approach offers a software route to improve the resolution-throughput trade-off in existing holotomography systems, allowing long-working-distance optics to recover laterally resolved RI features (e.g., subcellular puncta, internal texture) typically requiring short-working-distance high-NA optics.

Limitations and Scope
The authors explicitly note that the quantitative results are based on numerically emulated low-NA inputs generated from high-NA data. While this validates the inversion of the analytic operator, the method's performance on independently acquired physical low-NA data (which may contain unmodeled aberrations or noise) remains to be established. Furthermore, the current dataset is limited to three cell types under air-immersion conditions; extension to other modalities or sample types requires specific forward operators and calibration data. The method is presented as a measurement-anchored generative prior, where features in the recovered high-frequency band should be interpreted as prior-guided estimates rather than direct measurements.