Light sheet microscopy single-frame autofocusing method based on structured light illumination and deep learning

Abstract

The application discloses a light sheet microscope single-frame autofocusing method based on structured light illumination and deep learning, which comprises the following steps: shaping a light beam to form a composite multiple-depth and laterally staggered structured light illumination sample and keeping the sample fixed; moving a detection objective lens to shoot multiple image stacks and input the image stacks into a depth Fourier neural network for training; using the same depth composite structured light illumination sample to shoot a sample image and inputting the sample image into the trained network for out-of-focus prediction; and finally completing out-of-focus compensation according to the output result of the network. The application realizes the out-of-focus prediction of the light sheet microscope by using a single-frame image, solves the problem that multiple images need to be scanned in the traditional autofocusing scheme, and improves the long-time high-quality imaging capability of the system; the depth Fourier neural network is fast and high in precision in out-of-focus judgment, is beneficial to improving the imaging speed of the light sheet microscope, has a small number of required training sets and good network generalization, and is helpful to improving the availability of the network.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical microscopic imaging, in particular to a light sheet microscope single-frame autofocusing method based on structured light illumination and deep learning. BACKGROUND

[0002] Optical microscopes have been widely used in the field of life science research due to their non-invasive characteristics. Among them, light sheet fluorescence microscopes adopt a strategy of lateral illumination and wide-field detection, so they have good optical layer cutting ability and fast imaging ability, effectively avoiding the fluorescence excitation of samples outside the focal plane, which is conducive to reducing the phototoxicity and photobleaching of living cell samples, and is one of the important tools for long-term dynamic observation of living cells. For example, researchers can use light sheet fluorescence microscopes to achieve time-lapse imaging of living specimens for several hours to several days, such as zebrafish and fruit fly development. The best imaging quality is based on good focusing of the light sheet system. However, during long-term imaging, the system needs to constantly cope with possible defocusing problems due to various factors, such as changes in temperature, changes in the refractive index of the imaging buffer, and aberrations introduced by the sample itself.

[0003] In recent years, in order to ensure the stable imaging quality of light sheet microscopes during long-term imaging, researchers have proposed some adaptive autofocusing methods. For example, the excitation light sheet can be fixed and the detection objective lens can be moved to realize three-dimensional scanning of the sample, and the defocus amount of the light sheet and the detection objective lens can be identified by calculating the relevant parameters (DOI: 10.1038 / nbt.3708; 10.1101 / 222497) and corrected by using optical elements. However, this method needs to perform three-dimensional scanning on the sample to obtain multiple frames of sample information, causing low autofocusing efficiency. The method of realizing fast autofocusing by pupil segmentation image phase detection (DOI: 10.1038 / s41592-021-01208-1) introduces a beam splitter in the detection path, and the real-time detected fluorescence is introduced into an additional defocus detection optical path, which can realize real-time drift detection and defocus correction of the system through a single frame image. However, the additional hardware introduced by this system will cause photon loss, and also increase the complexity of the system. By using deep learning, researchers have realized the identification of system defocus amount through two frames of defocus images without additional detection hardware (DOI: 10.1364 / boe.427099). However, the existing autofocusing methods in light sheet fluorescence microscopes all need two or more images to predict the defocus amount, which undoubtedly increases the additional exposure of the sample during defocus detection, increases the risk of sample phototoxicity and photobleaching, and reduces the imaging speed of the system. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a light sheet microscope single-frame autofocusing method based on structured light illumination and deep learning, which can effectively improve the autofocusing efficiency of the light sheet microscope and improve the stability of long-term imaging of the system.

[0005] In order to achieve the above-mentioned purpose, the technical scheme provided by the present application is as follows:

[0006] According to a first aspect of the present application, a light sheet microscope single-frame autofocusing method based on structured light illumination and deep learning is provided, which comprises the following steps:

[0007] 1) shaping the light beam by a beam shaping element to form a structured light which is composed of multiple depths and is staggered in the lateral direction;

[0008] 2) illuminating the sample with the depth-combined structured light formed in step 1) and keeping it stationary, while moving the detection objective lens along the detection optical axis direction, and shooting a plurality of image stacks covering the set depth range;

[0009] After each image in the image stack is preprocessed, it is input into a deep Fourier neural network to train the network, and the preprocessing specifically includes: dividing each image into multiple sub-images, performing Fourier transform on each sub-image, dividing by the value of the zero-order peak for normalization, extracting the adjacent data groups of the high-order peaks to form tensors, and splicing the corrected multiple tensors together;

[0010] 3) illuminating the sample with the same depth-combined structured light to shoot a sample image, and inputting the preprocessed sample image into the trained deep Fourier neural network for defocus prediction;

[0011] 4) completing defocus compensation according to the defocus prediction result obtained in step 3).

[0012] Further, in step 1), the depth-combined structured light has N depths along the direction of the detection optical axis, where N>=2, and the number of depths N and the depth difference Dz can be adjusted according to actual experimental conditions.

[0013] Further, the structured lights at different axial depths are staggered in the lateral direction, i.e. the single-frame image is divided into N sub-images illuminated by structured lights at different depths in the field of view of the detection camera.

[0014] Further, in step 2), the real defocus amount of each image in the image stack data set for network training is a known quantity.

[0015] Further, in the preprocessing of step 2), according to the depth number N and the illumination range of the depth complex structured light, each image taken is divided into N sub-images, and each sub-image is subjected to Fourier transform and normalized by dividing by the value of the zero-order peak, a tensor composed of multiple intensity information adjacent to one of the high-order peaks provided by the structured light in each Fourier transform image is extracted, multiplied by the corresponding correction factor, and the N corrected tensors are spliced and input into the depth Fourier neural network.

[0016] Further, the tensor of each sub-image and the corresponding correction factor are calculated according to the following formula:

[0017] factor N =NR(z focus ±m·Dz) / R N (z focus )

[0018] wherein subscript N represents the Nth sub-image; m≥0 is an integer, and the value of m·Dz is equal to the distance between the Nth sub-image and the ideal focal plane corresponding to the illumination structured light; the function R(z) is defined as the ratio of the value of the high-order peak to the value of the zero-order peak in the Fourier transform image of each sub-image in the sub-image stack formed by all sub-images in the same lateral region as the depth position changes; the function NR(z) is defined as the normalized R(z) curve of any image stack; z focus is the position of the ideal focal plane in the image stack; and R N (z focus ) is the ratio of the high-order peak to the zero-order peak in the Fourier transform image of the Nth sub-image of the image at z focus .

[0019] Further, the depth Fourier neural network is composed of three fully connected layers, the neural units of each fully connected layer are activated by a ReLU activation function, and a Dropout layer is introduced after the second fully connected layer according to the need; in the training process, the average absolute error between the predicted defocus amount and the actual defocus amount is used to evaluate the network performance on the validation data set, and the training of the network is stopped until the average absolute error value is no longer optimized; the trained network finally outputs a scalar with positive and negative values, which is used to indicate the defocus amount and defocus direction of the microscope.

[0020] According to the second aspect of the present specification, an optical sheet microscopic system is provided, which is implemented by using the method of the first aspect, and the system is divided into three parts of an excitation light path, a detection light path and a computer:

[0021] The excitation light path comprises, in sequence, a laser, a collimating lens, a cylindrical lens group, a beam shaping system, a scanning galvanometer, a scanning lens, a tube lens, and an excitation objective; the laser is used to emit laser light; the collimating lens is used for collimating the laser light; the cylindrical lens group is used to shape the collimated circular Gaussian light beam into a collimated long strip-shaped light beam; the beam shaping system is used to shape the incident light beam into a depth complex structured light; the scanning galvanometer is used to scan the excitation light beam along the transverse direction to form a uniform illumination light sheet or along the detection optical axis direction to change the axial illumination position of the light sheet; the scanning lens and the tube lens are used to conjugate the light beam emitted from the scanning galvanometer to the entrance pupil of the excitation objective; and the excitation objective is used to converge the light beam onto the sample plane for sample illumination.

[0022] The detection light path comprises, in sequence, a detection objective orthogonal to the excitation objective, a detection tube lens, a filter, and a camera; the signal emitted by the sample is collected by the detection objective, focused by the detection tube lens, and filtered of stray light by the filter, and then imaged onto the camera.

[0023] The computer is used to control the shaping of the light beam, the scanning of the scanning galvanometer, the illumination of the sample, and the imaging of the camera, and is used for training of the depth Fourier neural network and subsequent defocus amount prediction of the system under single-frame imaging.

[0024] Further, the beam shaping system in the excitation light path is a related element capable of adjusting the exit amplitude distribution or phase distribution of the light beam, and the core element adopts an amplitude-type spatial light modulator, a phase-type spatial light modulator, or a digital micromirror array.

[0025] Further, the element for compensating the defocus amount adopts a scanning galvanometer along the detection optical axis direction conjugated to the entrance pupil surface of the excitation objective, a detection objective, or a detection tube lens, and the defocus amount compensation can be completed by adjusting the bias angle of the scanning galvanometer or the axial position of the detection objective or the detection tube lens.

[0026] The method provided by the application can realize defocus prediction of the light sheet microscope by using a single-frame image, solves the problem of needing to scan multiple frames of images in the traditional self-focusing scheme, reduces the number of additional exposures to the sample, helps to reduce the photo toxicity and photobleaching problem of the sample, and improves the long-time high-quality imaging capability of the system. In addition, the depth Fourier neural network has a fast defocus amount judgment speed and high precision, is beneficial to improving the imaging speed of the light sheet microscope, has a small number of required training sets and good network generalization, and is helpful to improving the availability and wide application of the network. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1Schematic diagram of depth composite structured light with different number of depths for embodiments of the present application, wherein (a) cross-sectional intensity distribution diagram of structured light illumination light sheet that composites two depths; (b) cross-sectional intensity distribution diagram of structured light illumination light sheet that composites three depths.

[0028] Figure 2 Schematic diagram of defocus prediction principle using depth composite structured light illumination for embodiments of the present application, wherein (a) image stack collected using structured light illumination sample that composites three depths; (b) Fourier transform image of any subgraph; (c) R N (z) curves, wherein N = 1, 2, 3; (d) each R N (z) curves are each normalized.

[0029] Figure 3 Structural schematic diagram of preprocessing module for embodiments of the present application.

[0030] Figure 4 Structural schematic diagram of depth Fourier neural network for embodiments of the present application.

[0031] Figure 5 Workflow diagram of depth Fourier neural network for embodiments of the present application, wherein (a) flowchart of network training; (b) flowchart of network work.

[0032] Figure 6 Schematic diagram of light sheet microscopy system based on depth composite structured light illumination for embodiments of the present application.

[0033] Figure 7 Effect diagram of number of training sets of depth Fourier neural network for embodiments of the present application.

[0034] Figure 8 Effect diagram of structural generalization of depth Fourier neural network for embodiments of the present application, wherein (a) effect diagram of defocus amount test of microtubule sample using neural network trained using microtubule sample; (b) effect diagram of defocus amount test of microfilament sample using neural network trained using microtubule sample. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described below in combination with embodiments and their accompanying drawings.

[0036] The present embodiment provides a light sheet microscope single-frame self-focusing method based on structured light illumination and deep learning, and the specific process is as follows:

[0037] 1) The light beam is shaped by a beam shaping element to form structured light that composites multiple depths and is staggered in the lateral direction;

[0038] 2) Using the depth compound structured light illumination sample formed in step 1) and keeping it fixed, while moving the detection objective along the detection optical axis direction, a plurality of image stacks covering a set depth range are taken;

[0039] Each image in the image stack is pre-processed and input into the depth Fourier neural network for network training, and the pre-processing is specifically: each image is segmented into a plurality of sub-images, each sub-image is subjected to Fourier transform and divided by the value of the zero-order peak for normalization, the adjacent data of the high-order peak is extracted to form a tensor, and the plurality of tensors are corrected and spliced together;

[0040] 3) Using the same depth compound structured light illumination sample, a sample image is taken, and the sample image is pre-processed and input into the depth Fourier neural network which has been trained to perform defocus prediction;

[0041] 4) Defocus compensation is completed according to the defocus prediction result obtained in step 3).

[0042] Specifically, in step 1), the depth compound structured light has N depths along the direction of the detection optical axis (here defined as the z-axis), where N≥2, which can be adjusted to a larger number of depths N and depth differences Dz according to actual experimental conditions; the structured lights at different axial depths are staggered in the transverse direction (here defined as the x-axis), that is, a single frame image is segmented into N sub-images illuminated by structured lights at different depths within the field of view of the detection camera, as shown in Figure 1 .

[0043] In step 2), taking the compound of three depths with a depth difference Dz of 0.5 microns as an example, using the depth compound structured light illumination sample as shown in Figure 1 (b) and keeping it fixed, while moving the detection objective along the detection optical axis direction, a plurality of image stacks covering a set depth range are taken, as shown in Figure 2 (a); in the image stack dataset for network training, the real defocus amount of each image is a known amount.

[0044] The pre-processing process of step 2) is realized by a pre-processing module, each image in the stack image is segmented into 3 sub-images, as shown in Figure 2 (a); each sub-image is subjected to Fourier transform and divided by the value of the zero-order peak for normalization, as shown in Figure 2 (b); 8×1 intensity information groups adjacent to one of the high-order peaks provided by the structured light in each Fourier transform image are extracted to form a tensor, and the corrected 3 tensors are spliced and input into the depth Fourier neural network, as shown in Figure 3 . The calculation formula of the correction coefficient corresponding to the 8×1 tensor in each sub-image is:

[0045] factor1 = NR(z) focus +Dz) / R1(z focus (1)

[0046] factor2=NR(z focus ) / R2(z focus (2)

[0047] factor3=NR(z focus -Dz) / R3(z focus (3)

[0048] Where the subscripts N = 1, 2, 3 refer to the Nth sub-image; the R(z) function is defined as the ratio of the value of the higher-order peak to the value of the zero-order peak in the Fourier transform image of each sub-image in the stack of sub-images formed by all the same horizontal regions, as a function of depth position. Figure 2 (b) and Figure 2 As shown in (c); the NR(z) function is defined as the normalized R(z) curve of any image stack, as shown in (c). Figure 2 As shown in (d); z focus Let R be the location of the ideal focal plane in the image stack; then R N (z focus ) for in z focus The ratio of higher-order peaks to zero-order peaks in the Fourier transform image of the Nth sub-image of the image at a given location.

[0049] like Figure 4 As shown, the deep Fourier neural network in this embodiment consists of three fully connected layers, containing 64, 32, and 32 neurons respectively. Each neuron is activated by the ReLU activation function. A Dropout layer is introduced after the second fully connected layer as needed. See [link to documentation]. Figure 5 In (a), the tensor information extracted by the preprocessing module from the acquired image stack is input into the deep Fourier neural network for training. During the training process, the mean absolute error between the predicted defocus amount and the actual defocus amount is used to evaluate the network performance on the validation dataset. The training of the network is stopped when the value of the mean absolute error can no longer be optimized. The trained network finally outputs a scalar with positive and negative values ​​to refer to the defocus amount and defocus direction of the microscope.

[0050] In step 3), as Figure 5 As shown in (b), the sample is illuminated using the same depth composite structured light, and a defocused image frame is acquired and input into the preprocessing module and a pre-trained deep Fourier neural network with appropriate weight parameters for defocus prediction. Finally, defocus compensation is completed based on the network output.

[0051] To achieve the above method, see Figure 6 The present application provides a light sheet microscopy system, which can be divided into three parts of excitation light path, detection light path and computer:

[0052] The excitation light path includes laser 1, collimating lens 2, cylindrical lens groups 3 and 4, polarization beam splitter 5, half-wave plate 6, spatial light modulator 7, first lens 8, annular mask plate 9, relay lens groups 10 and 11, x scanning galvanometer 12, scanning lens 13, tube lens 14 and excitation objective 15 placed in turn. The laser 1 is used to emit laser; the collimating lens 2 is used for collimation of laser; the cylindrical lens groups 3 and 4 are used to form a collimated long strip beam from a collimated circular Gaussian beam; the polarization beam splitter 5 is used to extract the s-polarized light of the light beam into the spatial light modulator 7; the half-wave plate 6 is used to optimize the polarization state of the light beam incident on the spatial light modulator 7; the spatial light modulator 7 is used to shape the light beam to form a depth complex structured light; the first lens 8 is used to Fourier transform the image emitted from the spatial light modulator 7; the annular mask plate 9 is used to filter out the zero-order and higher-order diffraction light emitted from the spatial light modulator 7, and only the ±1 order diffraction light is retained into the subsequent system; the relay lens groups 10 and 11 constitute a 4f imaging system, which conjugates the image emitted from the annular mask plate 9 to the x scanning galvanometer 12; the x scanning galvanometer 12 is used to quickly scan the excitation light sheet along the transverse direction, i.e. the x axis, to form a uniform light sheet to illuminate the sample; the scanning lens 13 and the tube lens 14 are used to conjugate the light beam emitted from the x scanning galvanometer 12 to the entrance pupil of the excitation objective 15, so that the annular mask plate 9 and the x scanning galvanometer 12 are conjugated to the entrance pupil of the excitation objective 15; the excitation objective 15 is used to converge the light beam to the sample plane to illuminate the sample.

[0053] The detection light path includes detection objective 16 orthogonal to the excitation objective, detection tube lens 17, optical filter 18 and camera 19 placed in turn. The signal emitted by the sample is collected by the detection objective 16, focused by the detection tube lens 17, filtered by the optical filter 18 to remove stray light, and then imaged onto the camera 9.

[0054] The computer 20 is used to control the operation of the spatial light modulator 7, the scanning of the x scanning galvanometer 12, the illumination of the sample and the imaging of the camera 19; and is used for training of the depth Fourier neural network and prediction of the defocus amount of the system in subsequent single-frame imaging.

[0055] Among them, the spatial light modulator 7 in the excitation light path is the core element of the light beam shaping system, which adjusts the exit amplitude distribution of the light beam, and other typical examples can be digital micromirror array.

[0056] In step 4), according to the output of the deep Fourier neural network, defocus compensation can be performed by adjusting the axial position of the detection objective 16. In addition, other defocus compensation schemes include but are not limited to adjusting the bias angle of the scanning galvanometer of the z-axis scanning light sheet conjugate to the entrance pupil plane of the excitation objective 15 or adjusting the axial position of the detection tube lens 17 to make the excitation light sheet coincide with the detection objective focal plane.

[0057] For testing effect, the system uses a detection objective with a numerical aperture of 1.1 for imaging, and uses a microtubule sample to collect a data set to train the deep Fourier neural network. As shown in FIG. 6, the neural network only needs 55 data sets to have good defocus prediction effect, that is, the root mean square error RMSE between the predicted value and the true value converges. Compared with the conventional neural network which usually needs more than 100 data sets, the deep Fourier neural network greatly reduces the demand for data quantity. Figure 7

[0058] As shown in FIG. 7, the deep Fourier neural network trained only with the microtubule data set has good defocus prediction effect on subsequent microtubule samples and microfilament samples, and has a lower RMSE within ±0.7 μm. As shown in FIG. 7(a), the prediction accuracy RMSE of the microtubule sample is 0.1024 μm, and as shown in FIG. 7(b), the prediction accuracy RMSE of the microfilament sample is 0.1140 μm, which shows that the neural network method of the present application has high prediction accuracy and good sample structure generalization. Figure 8 Figure 8 Figure 8

[0059] The above only describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.​​​​

Claims

1. A single-frame autofocusing method for light sheet microscopy based on structured light illumination and deep learning, characterized in that, The method includes the following steps: 1) The beam is shaped by a beam shaping element to form a structured light with multiple depths that are staggered in the lateral direction; 2) Illuminate the sample using the deep composite structured light formed in step 1) and keep it fixed, while moving the probe objective along the probe optical axis to capture multiple image stacks covering a set depth range. Each image in the image stack is preprocessed and then input into a deep Fourier neural network for training. The preprocessing specifically involves: dividing each captured image into N sub-images based on the depth N of the deep composite structured light and the illumination range; performing a Fourier transform on each sub-image and normalizing by dividing by the value of the zero-order peak; extracting multiple intensity information adjacent to one of the higher-order peaks provided by the structured light in each Fourier-transformed image to form a tensor; multiplying this tensor by the corresponding correction coefficient; and concatenating the corrected N tensors before inputting them into the deep Fourier neural network. The formula for calculating the correction coefficient for each sub-image's tensor is: factor... N =NR(z) focus ±m·Dz) / R N (z focus ), where the subscript N refers to the Nth subimage; m≥0 is an integer, and the value of m·Dz is equal to the distance between the illumination structure light corresponding to the Nth subimage and the ideal focal plane; the R(z) function is defined as the ratio function of the higher-order peak value to the zero-order peak value in the Fourier transform image of each subimage in the stack of subimages formed by all the same lateral regions, as a function of depth position; the NR(z) function is defined as the normalized R(z) curve of any image stack; z focus Let R be the location of the ideal focal plane in the image stack; then R N (z focus ) for in z focus The ratio of higher-order peaks to zero-order peaks in the Fourier transform image of the Nth sub-image of the image at the location; The deep Fourier neural network consists of three fully connected layers. The neurons in each fully connected layer are activated by the ReLU activation function, and a Dropout layer is introduced after the second fully connected layer as needed. During training, the mean absolute error between the predicted defocus amount and the actual defocus amount is used to evaluate the network performance on the validation dataset. Training stops when the mean absolute error value no longer needs optimization. The trained network finally outputs a scalar with positive and negative values ​​to represent the defocus amount and direction of the microscope. 3) Illuminate the sample using the same depth composite structured light, capture a frame of sample image, preprocess the sample image and input it into the pre-trained deep Fourier neural network for defocus prediction. 4) Complete the defocus compensation based on the defocus prediction result obtained in step 3).

2. The single-frame autofocusing method for light sheet microscopy based on structured light illumination and deep learning according to claim 1, characterized in that, In step 1), the depth composite structured light has N depths along the direction of the probe optical axis, where N≥2, and can be adjusted to more depths N and depth differences Dz according to actual experimental conditions.

3. The single-frame autofocusing method for light sheet microscopy based on structured light illumination and deep learning according to claim 1, characterized in that, Structured lights located at different axial depths are staggered in the lateral direction, which means that within the field of view of the detection camera, a single frame image is divided into N sub-images illuminated by structured lights at different depths.

4. The single-frame autofocusing method for light sheet microscopy based on structured light illumination and deep learning according to claim 1, characterized in that, In step 2), the true defocus amount of each image in the image stack dataset used for network training is a known amount.

5. A light-sheet microscopy system implemented using the method described in any one of claims 1-4, characterized in that, The system consists of three parts: the excitation optical path, the detection optical path, and the computer. The excitation optical path includes a laser, a collimating lens, a cylindrical lens group, a beam shaping system, a scanning galvanometer, a scanning lens, a tube mirror, and an excitation objective, arranged sequentially. The laser emits laser light; the collimating lens collimates the laser light; the cylindrical lens group shapes the collimated circular Gaussian beam into a collimated elongated beam; the beam shaping system shapes the incident beam into a depth-composite structured light; the scanning galvanometer scans the excitation beam along the transverse direction to form a uniform illumination plate or scans along the probe optical axis to change the axial illumination position of the plate; the scanning lens and tube mirror conjugate the beam emitted from the scanning galvanometer to the entrance pupil of the excitation objective; and the excitation objective focuses the beam onto the sample plane for sample illumination. The detection optical path includes a detection objective, a detection tube, a filter, and a camera, which are placed sequentially and orthogonal to the excitation objective; the signal emitted by the sample is collected by the detection objective, focused by the detection tube and filtered out by the filter to remove stray light, and then imaged onto the camera; The computer is used to control beam shaping, scanning mirror scanning, sample illumination, and camera imaging, and is also used for training deep Fourier neural networks and predicting the defocus amount of the system in subsequent single-frame imaging.

6. The light-sheet microscopy system according to claim 5, characterized in that, The beam shaping system in the excitation optical path is a related element that can adjust the output amplitude distribution or phase distribution of the beam. Its core element is an amplitude-type spatial light modulator, a phase-type spatial light modulator, or a digital micromirror array.

7. The light-sheet microscopy system according to claim 5, characterized in that, The element used to compensate for defocusing is a scanning galvanometer, a probe objective, or a probe tube mirror that scans along the probe optical axis and is conjugate to the entrance pupil surface of the excitation objective. Defocusing compensation can be achieved by adjusting the offset angle of the scanning galvanometer or the axial position of the probe objective or probe tube mirror.

Images