Non-invasive and noise-robust light focusing using confocal wavefront shaping

Abstract

Wavefront-shaping is a promising approach for imaging fluorescent targets deep inside scattering tissue despite strong aberrations. It enables focusing an incoming illumination into a single spot inside tissue, as well as correcting the outgoing light scattered from the tissue. Previously, wavefront shaping modulations have been successively estimated using feedback from strong fluorescent beads, which have been manually added to a sample. However, such algorithms do not generalize to neurons whose emission is orders of magnitude weaker. We suggest a wavefront shaping approach that works with a confocal modulation of both the illumination and imaging arms. Since the aberrations are corrected in the optics before the detector, the low photon budget is directed into a single sensor spot and detected with high signal-noise ratio. We derive a score function for modulation evaluation from mathematical principles, and successfully use it to image fluorescence neurons, despite scattering through thick tissue.

Fig. 1. System schematic.

An incoming laser illumination propagates through a layer of scattering tissue to excite a fluorescent neuron behind it. The emitted light is scattered again through the tissue toward a detector (main camera). In a conventional imaging system, the image of the neuron in the main camera is highly aberrated. Moreover, due to the weak emission, the image is very noisy. By adding two SLMs modulating the excitation and emission wavefronts, we can undo the tissue aberration. This allows us to reshape the incoming illumination into a spot on the neuron, and also to correct the emitted light and focus it into a spot in the main camera. For reference, we also place a validation camera behind the tissue, which can image the neuron directly and validate that the excitation light has focused into a spot. This camera does not provide any input to our algorithm. We visualize example images from these two cameras with and without the modulation correction.

Fig. 2. Types of fluorescent data.

a, b emission from invitrogen fluorescent microspheres (excitation/emission at 640/680 nm). A single bead is excited and the emitted light scatters through the tissue to generate a wide speckle pattern in (a). In (b) we use an aberration correction in the imaging arm so that the sensor measures a sharp spot. With such synthetic sources we can image a speckle pattern at high SNR, but this is not always the case with real biological samples. For example, (c, d) demonstrate fluorescent emission from EGFP neurons (excitation/emission at 488/508 nm), which is an order of magnitude weaker. In (c) a single fluorescent spot is excited, and the limited number of photons it emits are spread over multiple pixels. Noise is dominant, and an attempt to measure the variance of this image will evaluate the noise variance and not only the speckle variance. In (d) aberration correction is applied in the optics. As all photons are collected by a single pixel, SNR is drastically improved. Note that images (c, d) are taken under equal exposure and equal excitation power.

Fig. 3. Wavefront shaping results.

We visualize views from the validation and main cameras, each row demonstrates a different tissue sample. a-b The excitation light as viewed by the validation camera at the back of the tissue. Due to significant scattering, at the beginning of the algorithm when no modulation (mod.) is available, a wide speckle pattern is generated. After optimization, the modulated wavefront is brought into a single spot. c-d By placing a band-pass filter on the validation camera, we visualize the emitted light with and without the modulation correction. e-f Views of the emitted light at the main front camera with and without the modulation correction. Note that this is the only input used by our algorithm. Without modulation, light is scattered over a wide image area and the image is noisy. A sharp clean spot can be imaged when the limited number of photons are brought into a single sensor pixel. g By correcting the emission such that a single spot is excited and leaving the imaging path uncorrected, we can visualize the actual aberration of a single fluorescent point source. The top two examples used a thin brain layer behind parafilm, and the lower one is a thick brain slice.

Fig. 4. Convergence with noise.

We visualize images captured by the main camera at a few iterations of our algorithm. In the beginning a small number of photons are spread over multiple sensor pixels and the resulting image is very noisy. As the algorithm proceeds and a wavefront shaping modulation is recovered, the low number of photons is brought to a single sensor spot and the measured image has a higher SNR. SNR is calculated by capturing 40 images with the same modulation and evaluating their variance.

Fig. 5. Imaging a wide area.

We image a thin fluorescent brain slice behind a scattering layer. The top two images correct scattering through chicken breast tissue and the lower ones through parafilm. a Image of the neuron from the main camera with no correction, strong scattering is present and the neuron structure is lost. b Image with our modulation correction, the neuron shape as well as some of the axons are revealed. c A reference image of the same neuron, from the validation camera. The arrow marks a spot at which the optimization has converged. This spot is darker as it bleached during optimization.

Fig. 6. Imaging inside a thick fluorescent brain slice.

We image a wide 3D target inside a 400 μm thick fluorescent brain slice. a A confocal image of the neuron from the main camera with no correction, strong scattering is present and the neuron structure is lost. b A confocal image with our modulation correction, the neuron shape as well as some of the axons are revealed. c A reference image of the same neuron, from the validation camera. Due to the 3D spreading of the fluorescent components, the validation camera cannot always capture an aberration-free image of the target.