Background:
The advent of head-mounted mini-microscopes1 offers a unique opportunity to assess cerebral hemodynamics over unprecedented time-scales in the brains of freely behaving animals, e.g., spanning days and months during brain aging or from disease onset to therapy-driven recovery. Mini-microscopes often utilize intrinsic optical signals (IOS), a molecular contrast mechanism in which the absorption of light by erythrocyte-bound hemoglobin is used to generate high-resolution (e.g., 6 µm), wide field of view (e.g., 3×3 mm2) maps of the cerebrovasculature and changes in cerebral blood volume (CBV)2. IOS imaging in preclinical models that recapitulate human disease has led to critical insights, such as linking cortical spreading depression to brain hypoxia3 and quantifying hemodynamic dysfunction in Alzheimer’s disease4. Despite these advantages, IOS requires a substantial photon flux, e.g., >2000 photons per 0.1 seconds per pixel to create an image. Therefore, continuous IOS imaging could heat brain tissue, which limits its application to short-duration image acquisitions, e.g., for 15-30 minutes4.

Hypothesis/Objective:
Although IOS imaging under ultra-low light conditions, e.g., 10×-30× lower photon flux, could prevent tissue heating and permit continuous imaging, biological information becomes corrupted due to heavy imaging noise. To circumvent this technical hurdle, we hypothesize that deep learning (DL) can be applied to reconstruct and accurately extract cerebral hemodynamics from noise-corrupted IOS images acquired under ultra-low light conditions.

Methods:
IOS images of CBV changes in n=5 adult C57BL/6 mice were acquired via a miniature microscope5 mounted over the sensorimotor cortex. A ground truth (GT) dataset comprising ~17200 images/mouse, and a dynamic range of 0-255 digital numbers (DN) was used. Gaussian noise that approximated the Poisson distribution of light photons was introduced to each IOS image at progressively high standard deviations (σ) of 2, 4, 8, …,32, and 64 DN to simulate imaging noise at successively lower photon counts. Reconstruction of the GT was attempted by applying an image-guided filter (FILT) vs. DL models on noise-corrupted IOS images. The DL models, in order of increasing complexity, were an auto-encoder (AE), a U-Net, and a generative adversarial network (GAN)6. Data were split 80:20 into training and testing sets and trained for a maximum of 20 epochs. Each model was trained and evaluated per animal. Analyses are shown for images with σ=32 DN noise, which corresponds to a 32× reduction in light flux and a noise standard deviation of 1/8th the imaging dynamic range.

Results:
As shown in the representative example in Fig. 1a-j, noise that was characteristic of ultra-low light acquisition degraded vessel visibility (Fig. 1b-d), distorted intensity histograms (Fig. 1e-g), and compromised the spatial frequency content (Fig. 1h-j) of IOS images. However, as shown in Fig. 1n-p, all convolutional DL models (i.e., AE, U-Net, and GAN) successfully reconstructed the original IOS images (GT, Fig. 1k) from their noise-corrupted counterparts (GT+Noise, Fig. 1l). As shown in Fig. 1m, standard image filtering (FILT) could not achieve similar results. Fig. 1q-r quantifies the substantial DL-based improvement in peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM), two metrics that reflect image reconstruction accuracy. The panel in Fig. 1u shows an example GT, noise-corrupted, and reconstructed time series of CBV changes in the vessel highlighted in Fig. 1s-t. These data demonstrate the feasibility of DL-based retrieval of the CBV timeseries. Fig. 1v-z shows maps of the correlation coefficient between GT and noise-corrupted or DL-based CBV changes in the mouse sensorimotor cortex, showing how DL-based reconstructions achieved higher correlations across most vessels compared to standard image filtering.

Conclusion:
Our results highlight the feasibility of accurately retrieving cerebral hemodynamics from ultra-low light image acquisitions via deep learning.

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
Fig. 1: Deep learning enables the retrieval of cerebral hemodynamics from noise-corrupted data. (a) A representative image of intrinsic optical signals (IOS) cortex, indicating the cerebrovasculature supporting the mouse sensorimotor cortex. A 150×150-pixel region of interest (ROI) centered on a small blood vessel is highlighted via a black-colored box for subsequent analyses. (b) A magnified view of the ROI shown in (a). Using the IOS signals in this ROI as the ground truth (‘GT’), (c-d) one can impose additive white Gaussian noise (‘Noise’, c) to create an IOS image corrupted with noise (‘GT + Noise’, d). Here, noise is added to replicate the impact of ultra-low light acquisition on image quality. For illustration, the added noise followed a Gaussian distribution with zero mean and a standard deviation equal to 1/8th the maximum dynamic range of the GT image, i.e., 32 DN or digital numbers. (e-g) Corresponding histograms showing the probability density function (PDF) of pixel-wise IOS intensity distribution in the GT (e), noise (f), and GT+Noise (g) images. (h-j) Spatial power spectra corresponding to e, f, and g, respectively. (k-p) A representative IOS image (‘GT’, k), its noise-corrupted counterpart (‘GT+Noise’, noise σ=32 DN, l), and reconstructions via a standard, image-guided filter (‘FILT’, m) versus those via an Auto-Encoder (n), a U-Net (o), and a Generative Adversarial Network (‘GAN’, p). (q-r) Box-plots showing the distribution of image quality metrics, peak signal-to-noise ratio (PSNR, q), and structural similarity index measure (SSIM, r). Data are shown from a random collection of 100 images, each from n=5 animals. Significances are shown at p=0.05 w.r.t. the performance of the standard image filter via a two-way ANOVA test followed by post-hoc analysis. (s-t) A vessel region (t) marked on a representative IOS image (s). (u) Panel showing the time series of blood volume fluctuations in the vessel (‘VES’) extracted from the ground truth (‘GT’), GT+Noise (noise σ=32 DN), and their reconstructions via FILT, AE, U-Net, and GAN. (v-z) Spatial maps showing the correlation coefficient ‘r’ computed between the time series of blood volume changes at each vessel and their counterparts with ‘GT+Noise’, ‘FILT’, ‘AE’, ‘U-Net’, and ‘GAN’. ‘r’ is color-coded using a ‘jet’ color-scheme and mapped atop each surface vessel in the field of view. H=high, L=low.

Author
Janaka Senarathna
Johns Hopkins University School of Medicine