Methods for localizing signal sources in localization microscopy

Abstract

Methods in localization microscopy for localizing signal sources (3.1, 3.2), in which - at least once for each pixel (1) of a detector (2) used to detect detection radiation, values ​​of a pixel-specific error parameter are determined and assigned to the pixel (1) in question and stored in a calibration data set; - Image data of a sample is captured pixel by pixel as pixel values ​​and stored in an image data set; - based on the captured image data, a number of pixel (1) comprising origin areas (4) of signal sources (3.1,3.2) are identified; - a point distribution function is adapted to the pixel values ​​of the respective origin areas (4); and - based on the point distribution function, the respective signal source (3.1, 3.2) is localized within the relevant origin area (4); characterized in that - for each pixel (1) the pixel-specific error parameter is compared with a threshold value; - each pixel (1) that has a value of the pixel-specific error parameter greater than the threshold in the calibration data set; and - when adjusting the point distribution function, the marked pixels (1) in the image data set are either ignored or replaced by interpolation.

Description

[0001] The invention relates to a method according to the preamble of the main claim.

[0002] Several localization microscopy techniques are known from the state of the art. Prominent examples are methods for localizing individual signal sources, especially individual emitters, according to the techniques designated by the abbreviations "PALM" (photoactivated localization microscopy; e.g., WO 2006 / 127692 A2) and "STORM" (stochastic optical reconstruction microscopy; US 2008 / 0032414 A1). An overview of localization microscopy can be found, for example, in the articles by Klein et al., 2014 (Klein, T. et al., 2014; Eight years of single molecule localization microscopy. Histochemistry and cell biology 141: 561–575) and Babcock et al. 2012) (Babcock, H. et al., 2012; A high-density 3D localization algorithm for stochastic optical reconstruction microscopy. Optical Nanoscopy 1: 6).

[0003] All known localization microscopy techniques share the common feature of acquiring information for localizing signal sources (emitters, emitting molecules) from pixelated 2D area sensors, which serve as detectors or cameras. Since only a few hundred to a few thousand photons are available per localization step for individual signal sources such as individual emitting molecules, highly sensitive detectors are required.

[0004] Examples of such sensitive detectors are EMCCD sensors (electron multiplying charge coupled device). However, the underlying CCD architecture of EMCCD sensors, with its serial readout and amplification processes, results in limited pixel counts and / or limited readout speeds.

[0005] As an alternative, so-called "scientific CMOS" or sCMOS sensors (CMOS = complementary metal oxide sensor) have become established in recent years. They combine the "active pixel sensor" architecture made possible by CMOS technology with very low readout noise and high quantum efficiencies.

[0006] The advantages of sCMOS sensors over conventional EMCCD cameras include high frame rates, a large number of pixels, and smaller pixels. Furthermore, they do not produce excess noise because electron multiplication does not occur in sCMOS sensors. Ultimately, this leads to higher effective quantum efficiency. Therefore, sCMOS sensors are also well-suited for use in localization microscopy.

[0007] Potential inhomogeneities in the signal behavior of individual pixels can negatively impact the accuracy of localizing individual emitters as radiation sources. While such signal inhomogeneities might, at worst, lead to a distorted image in normal imaging, they can cause directional mislocalization during the localization step in localization microscopy. Such inhomogeneities can be caused, in particular, by the aforementioned "Active Pixel Sensor" architecture.

[0008] Patent US 9,769,399 B2 and the technical article by Huang et al., 2013 (Huang, F. et al., 2013; Video-rate nanoscopy using sCMOS camera-specific single molecule localization algorithms. Nature methods 10: 653-658) describe the use of parameterized models to model noise behavior.

[0009] In this process, pixel-dependent noise is modeled as a Gaussian distribution. Modeling parameters include, for example, the mean of the Gaussian distribution (offset), the variance of the distribution, and / or the gain factor per pixel.

[0010] An offset can be determined, for example, by capturing a number of dark frames. These dark frames allow the identification of signals generated by the pixel despite the absence of a detection beam (dark noise; erroneous signals). An average value can then be calculated from this data and used as an error parameter.

[0011] The gain can be determined by capturing and evaluating a number of images at different, but known, photon numbers or illuminance levels. Based on the captured image data, a distribution function of the pixel values ​​can be calculated, utilizing, for example, the sum of the photon-induced variance ("shot noise") and Gaussian variance.

[0012] A disadvantage of this approach to correcting noise behavior, however, is that individual pixels that deviate significantly from the respective model can lead to considerable errors. Examples of such pixels include so-called "blinkers," which can result in randomly outputted signals (Wang, X. et al. 2006; Random telegraph signals in CMOS image sensor pixels. Electron Devices Meeting; IEDM'06, International IEEE 2006).

[0013] The invention is based on the objective of proposing a method in localization microscopy for the localization of signal sources, with which the disadvantages occurring in the prior art are reduced.

[0014] The task is solved using a localization microscopy technique for locating signal sources. In this technique, for each pixel of a detector used to capture detection radiation, values ​​of a pixel-specific error parameter are determined at least once and stored in a calibration dataset, assigned to the corresponding pixel. Image data of a sample is acquired pixel by pixel as pixel values ​​in an image dataset. Based on this acquired image data, a number of pixel-spanning origin regions of signal sources are identified. A point distribution function (PSF) is fitted to the pixel values ​​of the pixels in the respective origin regions. Using the PSF, the origin of each signal source within the respective origin region is determined in 2D and / or 3D.

[0015] According to the invention, in the calibration data set, the pixel-specific error parameter for each pixel is compared with a previously defined threshold value. Each pixel exhibiting a pixel-specific error parameter value greater than the threshold value is marked in the calibration data set. When adjusting the PSF to the pixel values ​​of the respective origin area, all or some of the marked pixels in the image data set are either ignored or replaced by interpolation.

[0016] Additionally or alternatively, instead of the pixel-specific error parameter, a derived pixel-specific error parameter can be determined for each pixel using the calibration dataset, and / or an ADU (Analog-Digital Unit) histogram of a number of pixel values ​​can be created. The PSF (Performance Factor) is then adjusted based on the derived error parameter or the respective ADU histograms.

[0017] For the purposes of this description, a pixel is understood to be a detector element of a detector, for example, a detector element of an sCMOS sensor. The detector has a number of detector elements, which are arranged, in particular, in a two-dimensional matrix or an array.

[0018] The calibration dataset stores the pixel values ​​acquired during a calibration measurement, as well as optionally the respective pixel-specific and / or derived pixel-specific error parameters, assigned to the respective pixels. The image dataset contains the image data in the form of pixel values ​​for each pixel, which are captured during image acquisition, for example, as raw data in a PALM or STORM process. Each calibration dataset, or datasets derived from it, can be attached to the image dataset, for example, as metadata. In this way, for example, a pixel marked in the calibration dataset can also be marked in the image dataset.

[0019] Interpolation of pixel values ​​from the image dataset can be performed using known methods, for example, regression methods, estimations and / or the calculation of means or medians.

[0020] A pixel-specific error parameter is a signal emitted by the pixel in question that arises not from detected radiation, but from other environmental and / or component-related influences. Environmental factors can cause such erroneous signals due to the detector's temperature and temperature changes. Component-related factors can include defective areas in the detector's semiconductor material (Wang, X. et al. 2006; Random telegraph signals in CMOS image sensor pixels. Electron Devices Meeting; IEDM'06, International IEEE 2006). An erroneous signal from a pixel will transmit the detection of at least one photon, even though no photon was actually detected, or fewer photons were detected than the pixel's sensitivity would require.

[0021] Pixel-specific error parameters, hereinafter also referred to as error parameters, include, for example, the so-called offset, a variance of the pixel values, especially of a time series, and the gain.

[0022] The pixel-specific error parameters compared with the threshold may already have been corrected, for example, with regard to their offset and / or gain.

[0023] Derived error parameters are quantities derived from acquired pixel values. In particular, in one embodiment of a method according to the invention, a Gaussian function can be fitted to the pixel values ​​of a pixel. The width of the Gaussian function, denoted by sigma, can be used as a derived error parameter of the variance. This approach is advantageous compared to prior art methods in which the variance of a pixel is determined assuming a Poisson distribution. A variance deviating from the Poisson distribution can have a wide variety of physical causes, which are not known in individual cases, and therefore the distribution function cannot necessarily be described analytically. Nevertheless, assuming a Gaussian function, the width of the distribution function can be approximated, thus providing a better representation of the noise behavior for pixels that deviate from a Poisson distribution.Variances determined in this way can also be compared with a threshold value. Pixels whose variances exceed the threshold value are marked and subsequently replaced using interpolation within the image dataset or ignored entirely.

[0024] Emitting molecules (emitters) can act as signal sources, and their two-dimensional or three-dimensional position is referred to as the origin. The origin is determined using the photoelectric field width (PSF) within a defined origin area. In localization microscopy, a PSF typically has a width (full width at half maximum) of at least 9 (3x3) to 16 (e.g., 4x4) or 25 (5x5) pixels.

[0025] The core of the invention is the improved localization of signal sources. By ignoring or interpolating pixel values ​​whose error parameters exceed a predefined threshold, a shift in the origin of the signal source towards the pixel with the, for example, excessively high pixel value is effectively prevented. Furthermore, the specific signal behavior of a pixel can be used to correct erroneous signals.

[0026] According to the invention, two basic steps are carried out to achieve an improved consideration of the real noise behavior of each individual pixel.

[0027] As a first step, each pixel of the detector is calibrated. This calibration can be performed once, regularly, or repeated as needed, for example, if the detector's properties change over its lifetime. The calibration data is stored in a calibration data set.

[0028] Detector calibration can be performed by first disabling all specific filters of the detector platform (camera) that reduce or suppress unwanted noise and / or blinking during normal imaging use. The same applies to reducing or suppressing faulty signals from so-called hot pixels, i.e., pixels that permanently emit a faulty signal. Subsequently, a number of dark frames can be acquired, and error parameters and calibration data for each pixel can be determined based on their pixel values.

[0029] In the second step, the calibration data is used to account for pixel-specific noise behavior, correct the respective pixel values ​​of the image dataset, and avoid or at least reduce the disadvantages known in the prior art. The calibration data from the calibration dataset is appended to image datasets of measurements performed, for example, using the PALM or dSTORM method, or made available in another way. For example, instead of the calibration data itself, a link to the stored calibration data is written into the metadata.

[0030] One embodiment of the method specifically takes into account the fact that only a limited number of photons are available per signal source in localization microscopy. Even with a detector exhibiting ideal noise characteristics, the accuracy of the localization is therefore limited.

[0031] In the first step of the procedure, the number of erroneous signals each pixel of the detector would contribute to subsequent localization due to its individual noise characteristics is determined. Next, the number of photons and, consequently, the expected localization accuracy in a planned experiment are determined. Based on this, the tolerable localization inaccuracy is defined. This tolerance threshold serves as the threshold value.

[0032] Threshold values ​​can be set that should not be exceeded by the variances of the pixel values, in particular the variances of the pixel values ​​of a time series, or by the offset values ​​of the pixel values.

[0033] Pixels whose calibration data values ​​exceed the threshold during calibration are marked in the calibration dataset. Since the calibration dataset is associated with, and specifically appended to, the corresponding pixels of the image dataset are also marked. These marked pixel values ​​can then be interpolated and replaced during the localization step, for example, using the pixel values ​​of neighboring pixels in the image dataset. Alternatively, the pixel values ​​of the marked pixels can be ignored.

[0034] The threshold can be set generally or for each experiment individually. For a general setting, for example, a fluorophore that emits the fewest photons can be used (e.g., tdEOS for use with PALM; Wang et al. 2014; Characterization and development of photoactivatable fluorescent proteins for single-molecule-based superresolution imaging. PNAS 111: 8452–8457).

[0035] If the threshold is set for individual experiments or for groups of experiments, the distribution of the marked pixels may change as a result.

[0036] Additionally or alternatively, the threshold can be set depending on the wavelengths of an expected or known detection radiation, the intensity of the detection radiation and / or the temperature of the detector.

[0037] To prevent the accuracy of the localization from being negatively affected by a large number of marked pixels, in an advantageous embodiment of the method the permissible number of marked pixels per origin area is limited either absolutely or on average to a maximum number. For example, a maximum of two, preferably a maximum of one, pixels per origin area are permissible.

[0038] If the maximum permissible number is exceeded, the detector in question will not be used, or the localization data for that source area will be marked with a warning. Image data from that source area and / or the localization data may still be used, but their use for subsequent analyses should be cautious.

[0039] The pixel values ​​of the unmarked pixels in the image dataset can be corrected using the calibration dataset, for example with regard to offset and / or gain.

[0040] According to a further embodiment of the method, a derived error parameter is determined for each pixel based on the calibration data set. Variances of the pixel values ​​can be determined as derived error parameters for each pixel. Advantageously, this is done not under the assumption of a Poisson distribution, but rather under the assumption of a Gaussian distribution (see explanation above), deviating from the prior art.

[0041] A derived error parameter can also be a photon transfer curve, which can be used to correct or calibrate the gain. For this purpose, a number n of images are acquired under homogeneous illumination and varying illuminance (m values), and the mean values ​​Mm and variance Vm are determined for each pixel. The respective photon transfer curve is then created based on the parameters Mm and Vm (see also: Long, F. et al, 2014; Effects of fixed pattern noise on single molecule localization microscopy. Physical Chemistry Chemical Physics 16: 21586-21594).

[0042] Regarding the quality of the localization, it is advantageous if, in a further embodiment of the method, the measured noise behavior of each pixel is determined in the form of an ADU histogram or the corresponding data and stored in the calibration dataset for retrieval. For this purpose, each pixel is assigned its own set of ADU histograms for different illuminance levels. This provides a potentially multidimensional LUT (Lookup Table) for each pixel, from which an applied localization algorithm can read the actual noise behavior for each pixel, depending on the current local conditions. The localization is calculated based on the PSF (Pulse Frequency Function) taking into account the derived error parameter or the respective ADU histograms. For example, a pixel value is recorded as a grayscale value.Using the LUT, a real or actual gray value can be determined for the pixel in question and used for further processing. This approach makes it possible to determine the actual noise behavior of a given pixel even before the localization step and to perform the localization based on this actual noise behavior. This process can be applied to all pixels or only to selected pixels.

[0043] To ensure efficient processing and to achieve a compromise between resolution—and thus an accurate representation of pixel behavior—and the computational effort or the amount of metadata, the data from the respective ADU histograms can be appropriately combined (binned). Additionally, the ADU histograms can be smoothed and / or interpolated and / or adjusted using functions, and their parameters can be stored.

[0044] The derived error parameters represent the actual noise behavior of the pixels in question to a particularly high degree, which advantageously allows for improved modeling and more precise localization. In contrast to the approximation according to Huang et al., 2013 (Huang, F. et al., 2013; Video-rate nanoscopy using sCMOS camera-specific single molecule localization algorithms. Nature methods 10: 653-658), the actual noise behavior of the pixels is reproduced as accurately as possible. This advantageously opens up the possibility of better utilizing the detector's performance by, for example, comprehensively taking into account noise and blinking effects of the detector, which is hardly possible with prior art methods.

[0045] The embodiments of the method according to the invention advantageously avoid or reduce disadvantages known from the prior art. For example, high computing power is required for detector calibration and localization according to the prior art. In addition to the measurement and computational effort for calibration using, for example, 2.5 million frames (US 9,769,399 B2 and Huang et al. 2013; so), the noise of each pixel, as a convolution of shot noise (Poisson distribution) and pixel-dependent noise (Gaussian distribution), must be taken into account in the likelihood function of the "Maximum Likelihood Estimator" based on the calibration measurements. Since this must be done for each pixel during every localization process, the computational effort is very high. Therefore, this noise distribution is described by an analytical approximation, which, however, is only sufficiently accurate for those pixels that behave according to the assumed model.Pixels that deviate significantly from the model cannot be taken into account by adjusting the parameterized model and then lead to completely incorrect predictions by the model.

[0046] The invention is explained using the characteristics of sCMOS sensors as an example. However, it can also be applied to other types of pixel-based detectors. For example, the problems and effects described above also apply in principle to CCD and EMCCD sensors, although usually to a lesser extent.

[0047] The invention is explained in more detail below with reference to illustrations and various embodiments. These show: Fig. 1a an exemplary representation of captured pixel values ​​with a bad pixel and an emitting molecule as signal sources, and Fig. 1b an exemplary representation of a histogram of the determined PSF widths of the signal sources of Fig. 1a. Fig. 2 time series of gray value distributions of three different pixels of an sCMOS sensor under constant illumination; Fig. 3 a schematic representation of a method according to the invention using a threshold value; Fig. 4 a schematic representation of a calibration of the detector with various possible calibration parameters and a generation of a calibration data set; Fig. 5 a schematic representation of an exemplary implementation of pixel value processing and the calibration data set;

[0048] An exemplary representation of a two-dimensional arrangement of pixels 1 of a detector 2 shown in perspective, for example of an sCMOS sensor, is shown in Fig. Figure 1a is given. The different gray values ​​recorded by the individual pixels 1 are visible (three of them highlighted by an additional frame). Two particularly bright pixels 1 are selected as potential signal sources 3.1 and 3.2, and their origin areas 4 are visualized by means of a circle each.

[0049] In the Fig. Figure 2 shows an example of the temporal noise behavior of three different pixels 1 of an sCMOS sensor at constant illumination. The left column shows the gray values ​​plotted against time. The right column shows the corresponding histograms of the frequencies of the individual gray values.

[0050] The first line shows the noise behavior of pixel 1, which can be well described by a state-of-the-art model consisting of a convolution of the Poisson distribution and the Gaussian distribution.

[0051] The noise behavior shown in the last line cannot be adequately described by a corresponding model due to its large width and approximately triangular shape with a large base width.

[0052] The middle row shows the noise behavior of pixel 1, the distribution of which - as can be seen on the right in the corresponding histogram - cannot be captured using the known models.

[0053] One possible embodiment of the method according to the invention using a threshold value is described in Fig. Figure 3 is shown as a diagram. First, a threshold value to be applied is determined. This takes into account which experiment is to be performed. For example, properties of the sample and the emitters to be used, the illumination wavelengths and illuminance(s) to be used, the detection wavelength(s), as well as known specifications of the detector and / or the other experimental setup are considered.

[0054] Based on this, the expected localization accuracy and acceptable tolerances are determined or estimated. The threshold is then defined according to these considerations and specifications and stored, for example, in an evaluation unit 5. The threshold is defined, for instance, as an offset or as a variance of the pixel values, in order to compare it with the pixel-specific error parameters derived from those pixel values.

[0055] In an alternative embodiment of the procedure, the threshold can also be determined empirically using a localization algorithm.

[0056] In addition, a calibration data set is created and assigned to the individual pixels 1 of detector 2 (see Fig. 1a) assigned and stored. For this purpose, a number of dark frames are acquired with detector 2, and the recorded pixel values ​​are assigned to each pixel 1 and stored as a calibration data set. Subsequently, an offset or variance of the pixel values ​​is determined from the pixel values ​​as pixel-specific error parameters. These can also be stored in the calibration data set and assigned to pixels 1.

[0057] The values ​​of the pixel-specific error parameters determined in this way are compared with the threshold value for each pixel. If a pixel-specific error parameter exceeds the threshold value, the corresponding pixel is marked in the calibration dataset, for example, by listing its pixel coordinates (x, y) in the metadata. The marked pixels are to be disregarded for subsequent evaluations and are therefore "masked," i.e., marked as not to be considered further (pixel mask).

[0058] In a further step, it is checked whether masked pixel 1s are in clusters. For this purpose, the maximum permissible clustering is defined (specified; spec.) beforehand and also stored. For example, it is specified how far apart two marked pixel 1s must be at least to avoid forming a cluster.

[0059] If the conditions for permissible clustering are met, the process proceeds to the localization of the signal sources 3.1, 3.2.

[0060] If the distribution of the masked pixels 1 does not correspond to the permissible clustering, it is checked whether a permissible clustering can be achieved by changing the operating parameters (Spec.) of detector 2 (cluster specification). If this is not the case, then detector 2 (= sensor) is not suitable.

[0061] If, on the other hand, the operating parameters of detector 2 can be adjusted so that an acceptable clustering is achieved, these settings are made and the process proceeds to the localization of the signal sources 3.1, 3.2.

[0062] The calibration data set and the information on the marked pixel 1 (pixel mask) are provided as metadata and attached to an image data set to be created.

[0063] For a sample to be evaluated, a time series of image data is acquired (time series for localization microscopy), and the pixel values ​​are assigned to the respective pixels 1 and stored as an image data set. The masked pixels 1 are either assigned pixel values ​​determined by interpolation based on the pixel values ​​of neighboring pixels 1, or the masked pixels 1 receive no pixel value and are ignored during subsequent localization of the signal source 3.1, 3.2. Alternatively, it is also possible to assign the respective acquired pixel values ​​to the marked pixels 1, for example, to perform a separate error analysis. However, for actual localization, these pixels 1 are also ignored, or their pixel values ​​used for localization are determined and assigned beforehand by interpolation.

[0064] The pixel values ​​of the image dataset are corrected using the calibration dataset and, if necessary, taking the metadata into account (correction according to metadata). Based on the image data, especially the pixel values, origin regions 4 are determined within the array of pixels 1, in which a signal source 3.1, 3.2 is located. This is done, for example, using the maxima of the acquired and corrected pixel values. A localization algorithm is applied to the corrected pixel values ​​of pixels 1 in the respective origin region, and a PSF is adjusted. The location of the PSF maximum is determined as the origin of the signal source 3.1, 3.2 and stored.

[0065] The Fig. Figure 4 illustrates an example of detector calibration using various possible calibration datasets and the generation of a complete calibration dataset. For this purpose, time series with different illuminances I1 to Ik are acquired. The illuminance I1 is zero and corresponds to a time series of dark frames. The pixel values ​​acquired at illuminance I1 are used to generate a calibration dataset [Cal Interp (Mask)] with marked and masked pixels, as shown in Figure 4. Fig. 3 was explained.

[0066] The pixel values ​​acquired with the illuminance I1 can also be used to generate a calibration data set (Cal Offset), in which the offset of each pixel 1 is determined and stored. Pixel values ​​can be corrected by subtracting the mean of the pixel values ​​from each individual pixel value. To prevent negative pixel values ​​from occurring, or, depending on the data format, to avoid negative values ​​being prohibited (so-called "clipping" at zero), a constant and known value ("NoiseMargin") is optionally added to all pixel values ​​and stored.

[0067] Based on the individual time series, the mean pixel values ​​and the variances within the time series can be determined. These means and variances are used to create and adjust (photon transfer curve fitting) a photon transfer curve (PTC). The resulting photon transfer curve is stored in a calibration dataset (cal gain correction).

[0068] ADU histograms [Histogram (I1), Histogram (I2), ...; Histogram (Ik)] can also be generated and saved from the pixel values ​​of the time series. These can then optionally be combined (binned), smoothed, approximated, and / or filtered before being saved as calibration datasets [Cal (Histograms)].

[0069] The aforementioned calibration data sets can be combined into a single, complete calibration data set. Alternatively, they can also be created, saved, and / or used individually.

[0070] An application of the various calibration datasets to captured image data from an image dataset is exemplified in the Fig. Figure 5 shows that the image data, labeled T (Raw Data), is corrected for offset using the calibration data set Offset (Cal Offset) and for photon transfer using the calibration data set Gain (Cal Gain Correction). The pixel values ​​of the acquired image data, thus corrected, are then further corrected (masking) using the calibration data set Cal Interp (Mask). As explained above, those pixels in both the calibration data set Cal Interp (Mask) and the image data set whose variances exceed the threshold defined for the calibration data set Cal Interp (Mask) are marked and, if necessary, masked.

[0071] Based on the corrected pixel values, possible signal sources (Peak Finder) and associated origin areas 4 (Extract ROI; ROI = region of interest) are identified.

[0072] By adjusting the PSF [PSF-Fit (Gauss)], the origin of the signal source 3.1, 3.2 in the respective origin area 4 is determined and the quality of the adjustment is checked.

[0073] Additionally, after the actual localization, a filtering step can be performed. This allows the identification of "bad pixels" whose grayscale values ​​and noise behavior do not correspond to those of the emitter used in the experiment. Bad pixels include hot pixels, blinkers, dead pixels, or other types of pixel behavior that lead to apparent photon detection events. Such bad pixels may have been detected and localized as regular signal sources (3.1, 3.2), for example, as fluorescent molecules, despite the previously applied calibrations and localization steps.

[0074] Filter parameters can include, for example, the number of photons detected, the signal-to-noise ratio, and the blinking behavior. The width of the photon field (PSF) is also a suitable filter parameter, as it is largely independent of the specific experiment.

[0075] The latter filtering, in particular, relies on the fact that localizing individual signal sources requires a smearing of the PSF across multiple pixels 1, which occurs due to the optical design of the system, for example, the lens, tube lens, and pixel size. If significantly smaller PSF widths occur at individual pixels 1 localized by the algorithm, corresponding, for example, to a single pixel, these can be identified and excluded as bad pixels by appropriately selecting a filter threshold for the PSF width.

[0076] For example, in the Fig. Figure 1a shows two signal sources 3.1 and 3.2, each with its respective origin region 4 outlined in a circle. The pixel size 1 is 100 nm in the sample. The PSF widths (full width at half maximum, Gaussian fit) of the two signal sources 3.1 and 3.2 are shown in Fig. 1b shows an example for 1000 localizations each of both signal sources 3.1, 3.2. While the in Fig. The signal source 3.1 located in the upper left of 5a has a PSF width of only about 40 nm (range 30 - 50 nm), which is shown in Fig. 1a The signal source located at the bottom right 3.2 has a PSF width of about 140 nm, which is typical for a fluorophore.

[0077] It can therefore be assumed that the left signal source 3.1 is caused by a bad pixel (hot pixel or warm pixel or blinker), while the right signal source 3.2 is actually a fluorophore.

[0078] This filter function can, in principle, also be performed without the aforementioned steps of the inventive method. Reference sign 1 pixel 2 Detector 3.1 Signal source 3.2 Signal source 4 Origin area 5 evaluation unit

Claims

[1] Method in localization microscopy for localizing signal sources (3.1, 3.2), in which - at least once for each pixel (1) of a detector (2) used to detect detection radiation, values ​​of a pixel-specific error parameter are determined and assigned to the pixel (1) in question and stored in a calibration data set; - Image data of a sample is captured pixel by pixel as pixel values ​​and stored in an image data set; - based on the captured image data, a number of pixel (1) comprising origin areas (4) of signal sources (3.1,3.2) are identified; - a point distribution function is adapted to the pixel values ​​of the respective origin areas (4); and - based on the point distribution function, a localization of the respective signal source (3.1, 3.2) within the relevant origin area (4) is carried out; characterized by , that - for each pixel (1) the pixel-specific error parameter is compared with a threshold value; - each pixel (1) that has a value of the pixel-specific error parameter greater than the threshold in the calibration data set; and - when adjusting the point distribution function, the marked pixels (1) in the image data set are either ignored or replaced by interpolation. [2] Method according to claim 1, characterized by , that the threshold is determined depending on the wavelengths of a detection radiation, the intensity of the detection radiation and / or the temperature of the detector (2). [3] Method according to one of claims 1 and 2, characterized by , that on average a predetermined maximum number of pixels (1) is marked for each origin area (4). [4] Method according to any one of the preceding claims, characterized by, that the pixel values ​​of the unmarked pixels (1) are corrected using the calibration data set. [5] Method in localization microscopy for localizing signal sources (3.1, 3.2), in which - at least once for each pixel (1) of a detector (2) used to detect detection radiation, values ​​of a pixel-specific error parameter are determined and assigned to the pixel (1) in question and stored in a calibration data set; - Image data of a sample is captured pixel by pixel as pixel values ​​and stored in an image data set; - based on the captured image data, a number of pixel (1) comprising origin areas (4) of signal sources (3.1, 3.2) are identified; - a point distribution function is adapted to the pixel values ​​of the respective origin areas (4); and - based on the point distribution function, a localization of the respective signal source (3.1, 3.2) within the relevant origin area is carried out; characterized by , that - a derived error parameter is determined for each pixel (1) using the calibration data set and / or an ADU histogram of a number of pixel values ​​is created; and - the derived error parameter and / or the ADU histogram are stored in the calibration data set assigned to the respective pixels (1); - the adjustment of the point distribution function is carried out based on the derived error parameter or on the respective ADU histograms. [6] Method according to claim 5, characterized by, that as a derived pixel-specific error parameter, a variance of the pixel values ​​of a time series, a variance of the pixel values ​​of a time series at different illuminance levels and / or mean values ​​of a time series at different illuminance levels are determined. [7] Method according to claim 6, characterized by , that based on the variances from the time series of different illuminances and the mean values ​​a photon transfer curve for each pixel (1) is determined and used as a derived pixel-specific error parameter.

Images