Fig. 1. Principle of off-axis spectral encoding. (a) Schematic of off-axis intensity encoding under single-wavelength excitation. The detection and illumination axes of different line detectors have different off-axis distances. Line illumination beams with Gaussian distribution excite the sample, and each sub-detector of the multi-line detector records the corresponding modulated fluorescence signals at different off-axis positions. (b) PSF properties correspond to each sub-detector’s imaging in (a). (c) Schematic of off-axis spectral encoding under multi-wavelength excitation. Each sub-detector of the multi-line detector captures all fluorescence signals with distinct spectral mixing ratios due to wavelength-specific lateral offsets of each line-illumination beam. Different colors represent the excitation wavelengths of different channels. (d) PSF properties correspond to each sub-detector’s imaging in (c). (e) Modulation transfer function (MTF) curves for coaxial detection, off-axis detection, and diffraction limit. Scale bar: 500 nm.

Fig. 2. Principle of PSF linear decoding of off-axis spectral-encoded data. Ii represents the off-axis spectral-encoded raw image detected by Row i of the multi-line detector. Hi,j represents the PSF distribution of the jth wavelength channel in Row i of the multi-line detector. Xj represents the decoded jth wavelength channel image by the PSF linear decoding.

Fig. 3. Simulation of the PSF linear decoding using the virtual sample. (a) Off-axis spectral encoding data generation. The virtual sample containing four kinds of fluorescent labels is convolved with the 3D PSF of the off-axis spectral encoding to generate the raw mixed images of each sub-detector with serious crosstalk. (b) Single-wavelength-excited images of each kind of fluorescence label in the corresponding channels as the GT for the decoded results of the raw decoded images. (c) PSF linear-decoded results in four channels of the virtual samples with thicknesses of 0.5, 1, and 2 μm. Scale bar: 1 μm in (a), 20 μm in (b) and (c).

Fig. 4. Simulation of deep learning decoding for virtual sample. (a) Schematic diagram of the decoded network. (b) Deep-learning-decoded results of raw mixed images for virtual samples in Fig. 3(a) with thicknesses of 0.5 μm, 1 μm, 2 μm, and 10 μm from top to bottom. (c) Normalized intensities along the white lines in (b) and the same positions are shown in Fig. 3(c). Three peaks indicated by the three arrows correspond to the crosstalk from the three respective color channels. (d) Comparison of time consumption between PSF linear decoding and deep learning decoding. Scale bar: 20 μm.

Fig. 5. Schematic of the four-color off-axis spectral encoding line scanning microscope. The optical path mainly consists of three parts: beam combination and expansion, independent adjustment of the positions for each beam, and detection of multicolor fluorescence signals.

Fig. 6. Multicolor imaging of 200-nm-diameter fluorescent beads. (a) The raw-mixed (Raw, left), linear-decoded (Decoded, middle), and ground-truth (GT, right) images. The four rows of images are the images of three detection channels and their merged image. 488 nm, 561 nm, and 633 nm correspond to the excitation wavelength of each channel, and the corresponding decoded results are shown in cyan, yellow, and purple, respectively. The white boxed area shows the typical signal distribution of the three types of beads. (b) The line profiles and corresponding Gaussian fittings through the beads along a 45° diagonal from the images in (a). (c) The numbers of the fluorescent beads in the raw-mixed, linear-decoded, and GT images. Scale bar: 10 μm.

Fig. 7. Multicolor imaging of the mixture of four E. coli strains with different fluorescent labels. (a) PSF linear-decoded image of the mixture of four E. coli strains labeled by EBFP (cyan), EGFP (green), mScarlet (yellow), and smURFP (red). (b) Zoomed merged images of the white box in (a) using raw-mixed, linear-decoded, and ground-truth data. (c) Correlation between the proportion of each E. coli species detected from the PSF linear-decoded and the GT images. Data points of different colors and shapes correspond to different types of E. coli. (d) Comparing the data acquisition time of the GT images by four-color sequential imaging and those of the decoded images of off-axis spectral encoding four-color simultaneous imaging. Scale bar: 100 μm (a), 20 μm (b).

Fig. 8. Multicolor imaging of a 5-μm-thick BALB/c mouse brain slice. (a) Coronal image of the brain slice reconstructed by deep learning decoding, which contains four fluorescence signals of the nucleus (blue), neuron cytoplasm (green), neuron nucleus (magenta), and astrocyte (red). (b) Zoomed images of the yellow rectangular box in (a), showing the raw-mixed, PSF linear-decoded, deep-learning-decoded, and GT images from left to right. (c) Statistical results of structural similarity (SSIM) of the raw-mixed, linear-decoded, and deep-learning-decoded images with GT images. (d) Zoomed images of the square in (b). From left to right are the raw-mixed, linear-decoded, wide-field, line confocal, deep-learning-decoded, and GT images. The merged images and the decoded images of 405-, 488-, 561-, and 633-nm-excitation channels are from top to bottom. (e) Signal intensity profiles along the corresponding white lines in (d). Two black arrows indicate a bright and a weak protrusion. The black box indicates another weak protrusion, as indicated by a white arrow in (d). Scale bar: 1 mm (a), 100 μm (b), 15 μm (d).