Photon-Resolved Excitation-Denoised (PRED) Three-Photon Imaging Improves Detection of Neuronal Activity in Awake and Behaving Mice

The authors present Photon-Resolved Excitation-Denoised (PRED) three-photon imaging, a novel technique that overcomes previous limitations in speed, signal-to-noise ratio, and motion artifacts to enable high-speed, high-fidelity functional imaging of neuronal activity in deep brain regions of awake, behaving mice.

The Gist

Imagine trying to take a high-speed, crystal-clear photo of a tiny, glowing firefly deep inside a dense, foggy forest. That is essentially what neuroscientists are trying to do when they look at brain cells deep inside a living mouse's head.

For years, scientists had a powerful camera called Two-Photon Microscopy. It was great for seeing fireflies near the surface of the forest (the top layers of the brain). But as they tried to look deeper, the fog (brain tissue) got thicker, the light got scattered, and the images became blurry or too dim to see anything.

To see deeper, they invented Three-Photon Microscopy. This is like using a super-bright, specialized flashlight that can punch through the fog better. However, this new flashlight had its own problems:

1. It was too slow: To get a clear picture deep down, you had to move the flashlight very slowly, like a snail. But brain activity happens in the blink of an eye.
2. It was shaky: The flashlight flickered slightly, making the fireflies look like they were blinking when they weren't.
3. It was too hot: If you turned it up too high to see deeper, you risked burning the forest (damaging the brain).

This paper introduces a new system called PRED-3P that fixes all these problems, allowing scientists to take fast, clear movies of the brain's "deep forest" while the mouse is running around.

Here is how they did it, using some simple analogies:

1. The "High-Speed Train" Scanner (Speed)

Previously, the camera scanner was like a slow, old-fashioned train that could only stop at a few stations (pixels) per second. To see a whole city (a large area of the brain), it took forever.

• The Fix: They swapped the old train for a high-speed bullet train (a resonant scanner) that zips back and forth 8,000 times a second.
• The Catch: A bullet train moves so fast that if you only have one train car (laser pulse) per second, you can't stop at every station.
• The Solution: They used a laser that fires pulses like a machine gun (4 million times a second). Now, even though the scanner is moving at supersonic speeds, there are enough laser "bullets" to hit every single pixel. This allows them to film the brain at 20–30 frames per second—fast enough to catch the brain thinking in real-time.

2. The "Noise-Canceling Headphones" (The PRED System)

Even with a fast camera, the signal from deep in the brain is very weak. It's like trying to hear a whisper in a room full of static noise.

• The Problem: The laser itself flickers slightly (like a dimmer switch that jitters), and the camera adds its own electronic "hiss." Because the brain signal is so faint, these tiny errors make it look like the brain is active when it's actually quiet.
• The Fix (PRED): They built a system that acts like noise-canceling headphones.
  • They added a second sensor that measures the exact strength of every single laser pulse as it happens.
  • They used a super-sensitive detector (a Silicon Photomultiplier) that is cooled down to near absolute zero to stop it from "shivering" (electronic noise).
  • They used a smart computer algorithm (Bayesian statistics) that looks at the laser's strength and the camera's reading, then says, "Ah, the laser was 5% brighter this time, so I'll subtract that extra brightness from the image."
  • Result: They stripped away the "static," leaving only the true "whisper" of the brain cells.

3. The "Perfect Lens" (Beam Shaping)

When you shine a flashlight through fog, the edges of the beam scatter more than the center.

• The Fix: They carefully adjusted the shape of the laser beam (like adjusting the focus ring on a camera lens) to make sure the light was perfectly tight and concentrated exactly where they wanted it, without wasting energy on the edges. This meant they could see deeper without needing to turn up the heat (power) and burn the tissue.

The Result: Seeing the "Deep Brain"

With this new setup, the scientists were able to look at the Dentate Gyrus, a part of the mouse's hippocampus (the brain's GPS system) that is buried about 600–1,000 micrometers deep.

• Before: They could only see the top layers. The deep layers were a blur.
• Now: They can see hundreds of individual neurons firing while the mouse runs on a treadmill.
• The Discovery: They confirmed that deep brain cells have specific "addresses" (spatial tuning). Just like a neuron in the top layer knows "I am at the finish line," a neuron in the deep layer also knows exactly where the mouse is in the room.

Why This Matters

Think of the brain as a massive city. For years, scientists could only map the streets on the surface. They knew the city existed, but they couldn't see what was happening in the subway tunnels or the skyscrapers 50 floors down.

This new PRED-3P technology is like building a high-speed, noise-canceling, deep-penetrating subway camera. It allows us to finally map the "deep brain" while the animal is awake and behaving, opening the door to understanding how our deepest memories and thoughts are formed without having to cut open the brain or damage it with heat.

Technical Summary

1. The Problem

Three-photon microscopy (3PM) has revolutionized deep-brain imaging, allowing optical access to neurons ~500–1500 µm below the brain surface (e.g., the hippocampal dentate gyrus). However, its application in awake, behaving animals has been severely limited by three main bottlenecks:

• Speed vs. Field of View (FOV) Trade-off: 3PM requires high peak power but low average power to avoid thermal damage. To maximize efficiency, systems typically use low repetition rate lasers (~0.5–2 MHz) with one pulse per pixel. This limits imaging frame rates (often <10 Hz), which is too slow to capture the fast dynamics of modern calcium indicators or to cover large FOVs at high resolution.
• Signal-to-Noise Ratio (SNR) and Noise Sources: Deep imaging suffers from dim signals due to scattering and absorption. The signal is further corrupted by:
  • Laser Fluctuations: Because 3P excitation scales with the cube of pulse energy, minor laser power fluctuations cause massive brightness variations (e.g., a 5% power fluctuation results in a ~16% signal change).
  • Detector Noise: Standard detectors introduce electronic noise and dark counts that obscure the weak photon fluxes from deep tissue.
• Inaccessibility of Deep Sub-regions: While 3PM can reach deep structures, specific regions like the infrapyramidal blade (IPB) of the dentate gyrus and the hilus have remained largely inaccessible for functional imaging during behavior due to the depth and the inability to image fast enough to capture neuronal dynamics.

2. Methodology

The authors developed a comprehensive system integrating hardware optimization, novel detection strategies, and advanced statistical processing to overcome these limitations.

A. Hardware Optimization

• Laser & Scanning: They utilized a 4.14 MHz repetition rate laser (1300 nm) coupled with a resonant-galvo-galvo (resonant-GG) scanning engine capable of 8 kHz line rates. This combination allows for high-speed scanning (20–30 Hz) over a large FOV (~250 µm × 350 µm) while maintaining sufficient peak power per pulse for 3P excitation without exceeding thermal damage thresholds.
• Spatiotemporal Pulse Shaping:
  • Spatial: They empirically tested objective backfilling ratios ($\beta$). They found that a $\beta$ of 0.7 (slightly underfilled) provided the optimal balance between resolution and signal collection depth, outperforming both fully backfilled ($\beta=0.9$) and highly compressed ($\beta=0.6$) beams in deep tissue.
  • Temporal: They optimized pulse duration by using chirped mirrors for negative dispersion and variable thickness SF11 glass for positive dispersion compensation, ensuring the shortest possible pulses at the focal point.
• Detection: Replaced standard Photomultiplier Tubes (PMTs) with a deeply cooled Silicon Photomultiplier (SiPM) array. This detector offers high linearity and low dark counts, crucial for single-photon resolution.

B. Photon-Resolved Excitation-Denoised (PRED) Correction

This is the core algorithmic contribution. The system simultaneously records:

1. Fluorescence Signal: Detected by the SiPM.
2. Excitation Power: Measured pulse-by-pulse using an InGaAs photodiode sampling the laser beam.

The PRED algorithm uses Bayesian statistics to:

• Model Detector Response: Fit Gaussian templates for 0, 1, and 2-photon events to distinguish true signals from dark counts and crosstalk.
• Decouple Excitation Noise: Since the laser power is measured for every pixel, the algorithm infers the "true" photon count by normalizing the detected signal against the instantaneous laser power. It effectively removes the non-linear amplification of laser noise (the cubic dependence).
• Result: The output is a discrete photon-count estimate that approaches the Poisson limit, eliminating excess noise from electronics and laser instability.

C. Experimental Setup

• Subjects: Awake, head-fixed mice running on a cued burlap belt.
• Target: Dorsal hippocampus, specifically the Dentate Gyrus (DG) including the Suprapyramidal Blade (SPB), Hilus (HIL), and Infrapyramidal Blade (IPB).
• Indicator: Genetically encoded calcium indicators (jGCaMP8s or jGCaMP7f) expressed via AAV.

3. Key Contributions

1. High-Speed Deep Imaging: Achieved 20–30 Hz imaging of neuronal populations in the deep dentate gyrus (>600 µm depth) in behaving mice, a significant leap from previous 3PM limits of ~5–10 Hz.
2. PRED Algorithm: Introduced a robust method to eliminate excess noise (laser fluctuations and detector noise) by combining single-photon resolution with excitation power monitoring, effectively recovering the signal-to-noise ratio to the theoretical Poisson limit.
3. First Functional Imaging of IPB: Successfully imaged the infrapyramidal blade (IPB) of the dentate gyrus for the first time in vivo, a region previously inaccessible to 2P and difficult for 3P due to depth.
4. Optimized Pulse Shaping: Provided empirical data on optimal objective backfilling ($\beta=0.7$) and dispersion compensation for deep 3P imaging, challenging the assumption that full backfilling is always optimal.

4. Results

• Cell Detection: The PRED-3P system detected significantly more cells than 2P microscopy at all depths, with the advantage increasing as depth increased. In the IPB, 3P detected clear cellular structures where 2P failed.
• Noise Reduction: PRED correction reduced the noise factor to near 1 (Poisson limit) and eliminated spatial artifacts caused by laser ringing and scan-direction fluctuations. Simulations showed a dramatic increase in the ability to detect small ($\Delta F/F_0 \approx 20\%$) calcium transients.
• Functional Characterization:
  • Granule Cells (GCs) & Mossy Cells (MCs): The system successfully recorded calcium transients in both cell types during running.
  • Behavioral Correlation: Both GCs and MCs showed higher activity during running vs. rest.
  • Spatial Tuning: The authors identified spatially tuned Mossy Cells in the hilus, a finding consistent with previous literature but achieved here with a much larger FOV and higher speed. They observed that tuned MCs often have multiple place fields and fire larger transients than GCs.
  • Population Dynamics: Population-level analysis confirmed no specific reward-zone bias in transient rates, matching established physiological models.

5. Significance

This work represents a major breakthrough in deep-brain functional imaging:

• Overcoming the Depth-Speed Barrier: It proves that 3PM can be used for high-speed, large-FOV imaging in behaving animals, making it a viable alternative to 2PM for deep structures.
• New Biological Insights: By accessing the IPB and HIL, researchers can now study the functional heterogeneity of the dentate gyrus (e.g., differences between SPB and IPB granule cells) in the context of behavior, which was previously impossible.
• Generalizable Denoising: The PRED correction method is applicable to any low-photon-flux imaging scenario (e.g., voltage imaging or single-molecule microscopy) where laser stability and detector noise are limiting factors.
• Future Directions: The authors suggest that combining this approach with adaptive optics and specialized collection optics could further push the boundaries of deep-brain imaging, potentially enabling functional studies of cortical layers V & VI or other deep subcortical structures.