Supplementary Figure 9: Computational image reconstructions using image sub-samples.
Assessments of the potential improvements in imaging depth attainable by operating the high-speed two-photon microscope in a sub-sampling mode, in which the camera acquires multiple image frames during each scanning cycle of the galvanometer, such that scattered fluorescence photons can be computationally re-assigned to their most likely positions of origin in the specimen by using the known positions of the laser foci for each acquired image frame. Due to the speed limitations of existing scientific-grade cameras, these assessments were low-speed, proof-of-concept studies using image sub-samples acquired at 256 Hz, as well as a 24-µm-separation between the foci of adjacent laser beamlets in the specimen plane, a 200-kHz laser repetition rate, and fixed brain slices with fluorescent neurons expressing either tdTomato, YFP or GCaMP6 (Methods). The power per beamlet delivered to the sample ranged from 0.2 mW for images acquired ~100 μm deep within tissue, to 1 mW for images acquired ~500 μm deep within tissue. (a) Schematic illustration of the standard process of image formation used elsewhere throughout the paper. As the grid of laser beamlets sweeps across the image plane, the sCMOS camera sums the fluorescence signals captured across an entire laser-scanning cycle. In principle, however, a camera operating at faster image acquisition rates could capture the distinct fluorescence emission patterns for different positions of the laser beamlet grid. Access to these patterns permits superior methods of image reconstruction that correct for fluorescence scattering, as illustrated in panels b–d. (b) Schematic illustration of Reconstruction Method 1, in which a blind deconvolution algorithm is applied to each image sub-sample, followed by a maximum projection operation across the entire set of image sub-samples, to attain the final reconstructed image (Methods). The kernel for the deconvolution approach is initialized using the Ps(x,y) functions determined in Supplementary Fig. 8a for the relevant fluorophore and depth in tissue. (c) Schematic illustration of Reconstruction Method 2, which incorporates the knowledge that the laser illumination pattern has the form of a square grid. The orientation of the grid is determined for each image sub-sample by using a Radon transform, and then a peak-finding algorithm identifies the grid nodes (Methods). After this determination of grid geometry, within each image tile the fluorescence signals are re-assigned to the center of the tile. Next, to computationally correct for fluorescence scattering across nearby image tiles, the fluorescence signals assigned to each image tile undergo a blind un-mixing, using an un-mixing matrix that is initialized with the pi,j matrices determined in Supplementary Fig. 8d. After un-mixing, the resulting signals are re-mapped from each image tile back into a representation of the specimen plane. After this procedure is applied to each image sub-sample, the entire image is reconstructed by summing the results obtained from all the sub-samples. (d) Example images processed using each of the methods outlined in panels a–c, for cells expressing tdTomato imaged in fixed brain slices at three different depths in tissue. Reconstruction Methods 1 and 2 improved image contrast by 310 ± 130% (s.d.) and 310 ± 190%, respectively, as determined on a set of 59 different images acquired in fixed brain tissue slices with neurons expressing tdTomato, YFP or GCaMP6f, across a range of tissue depths from 0–615 μm beneath the tissue surface. Scale bars: 50 μm in a–c and 25 μm in d.
