AI Phase Retrieval in Fourier Ptychographic Microscopy

This article explains how Fourier Ptychographic Microscopy (FPM) overcomes traditional optical limitations to generate gigapixel pathology images. By combining a programmable LED array with artificial intelligence-driven phase retrieval algorithms, FPM captures multiple low-resolution images from varying illumination angles and computationally stitches them in the frequency domain. This process yields ultra-high-resolution, wide-field-of-view images containing both amplitude and phase information, eliminating the need for mechanical scanning in digital pathology.

Overcoming the Space-Bandwidth Product Limit

Traditional optical microscopes face a fundamental physical trade-off known as the space-bandwidth product limit. A microscope can either capture a wide field of view (FOV) at low resolution using a low-magnification objective, or a tiny FOV at high resolution using a high-magnification objective. Pathologists historically bypassed this by mechanically moving a tissue slide under a high-magnification lens and stitching hundreds of spatial images together.

Fourier Ptychographic Microscopy bypasses mechanical scanning entirely. It uses a low-magnification objective lens to immediately capture a wide FOV. To recover the high-resolution details lost by the low-magnification lens, FPM replaces the standard microscope light source with a two-dimensional array of Light Emitting Diodes (LEDs).

Angular Illumination and the Fourier Space

When light passes through a pathology slide, it diffracts. High-resolution, microscopic details diffract light at very wide angles—angles too steep for a low-magnification lens to collect.

FPM solves this by turning on individual LEDs in the array sequentially. Each LED illuminates the specimen from a distinct, known angle:

  • Central LEDs: Illuminate the sample directly, allowing the lens to capture low-spatial-frequency information (basic shapes and contrast), resulting in a standard low-resolution brightfield image.
  • Off-axis LEDs: Illuminate the sample at steep angles. This shifts the high-frequency diffracted light back into the acceptance cone of the objective lens. The camera captures these as darkfield images, which contain the ultra-fine structural details of the tissue.

According to optics physics, shifting the angle of illumination in physical space corresponds to shifting the captured data region in the spatial frequency domain, known as Fourier space. Each low-resolution image captured under a different LED represents a different circular aperture within a much larger Fourier space.

Stitching in the Frequency Domain

Instead of stitching images edge-to-edge in physical space, FPM stitches them together in the Fourier domain. The LEDs are spaced closely enough that the frequency data captured by one LED partially overlaps with the data captured by its neighboring LED. Typically, a 60% to 70% overlap in the Fourier domain is required.

By combining the low-frequency data from the central LEDs with the high-frequency data from the outer LEDs, the system synthesizes a massive synthetic aperture. This synthetic aperture mimics the light-gathering power (Numerical Aperture) of a high-magnification lens while retaining the wide field of view of the low-magnification lens.

Phase-Retrieval Artificial Intelligence

The core challenge of FPM is that digital camera sensors can only measure the intensity (amplitude) of light, completely losing the phase information (how much the light wave is delayed as it passes through varying tissue thicknesses). Without phase data, the overlapping regions in the Fourier domain cannot be accurately combined.

Traditionally, iterative phase-retrieval algorithms switched back and forth between the spatial domain and the Fourier domain to guess the missing phase data. Today, this bottleneck is resolved using phase-retrieval artificial intelligence and deep learning frameworks.

AI integrates into the FPM reconstruction pipeline in two primary ways:

  1. Physics-Informed Neural Networks: The FPM imaging process is modeled as a neural network. The unknown high-resolution complex image (containing both amplitude and phase) acts as the network's weights. The AI forward-models the low-resolution states, compares them to the actual captured LED images, and uses backpropagation and gradient descent to rapidly update the weights until the exact phase and amplitude are recovered.
  2. Deep Learning Reconstruction: Convolutional Neural Networks (CNNs) are trained on vast datasets of pathology slides. Once trained, these networks can take a sparse set of raw LED images and instantly infer the final high-resolution complex image in a single pass. This reduces computation time from hours to milliseconds and requires fewer LED illuminations, speeding up image acquisition.

Creating the Gigapixel Pathology Image

Once the AI completes the phase-retrieval and frequency-stitching process, an Inverse Fourier Transform is applied to the massive synthetic aperture. This mathematical step converts the consolidated frequency data back into a standard visual image.

The result is a gigapixel-scale pathology image. Because the AI successfully recovers both parts of the light wave, the output provides two distinct views from a single scan: a ultra-high-resolution color intensity image for standard histological analysis, and a quantitative phase map. The phase map reveals the optical thickness and refractive index of unstained cells, giving pathologists unprecedented structural data without the need for chemical staining or mechanical slide scanning.