Spectroscopic imaging method

Abstract

The method of imaging the at least one sample comprises providing an illumination light beam comprising light of one or more illumination wavelengths for illuminating the at least one sample. The method further comprises imaging the at least one sample at a plurality of first imaging positions. The imaging comprises illuminating, at the plurality of first illumination positions, the at least one sample with the illumination light beam; collecting first detection light from the plurality of first illumination positions; and generating, from the collection detection light, first imaging data associated with the plurality of first imaging positions, wherein, for at least one of the plurality first imaging positions, the first imaging data is generated from several ones of the plurality of first illumination positions. The method further comprises analysing the first imaging data with respect to at least one detection wavelength, to identify at least one region of interest (ROI). The method further comprises imaging the at least one ROI at a plurality of second imaging positions. The imaging of the at least one ROI at a plurality of second imaging positions comprises illuminating, at the plurality of second illumination positions, the at least one sample with the illumination light beam; collecting second detection light from the plurality of second illumination position; and generating second imaging data associated with the plurality of second positions. The method further comprises analysing the second imaging data with respect to the at least one detection wavelength.

Description

-1- DescriptionTitle: Spectroscopic imaging methodCROSS-REFERENCE TO OTHER APPLICATIONS

[0001] The present invention claims priority to the Luxembourg patent application LU508138 filed on 30 August 2024, the entire disclosure of which is incorporated herein by reference.FIELD OF THE INVENTION

[0002] The invention relates to a method and system for imaging samples, in particular cellular or tissue structures.BACKGROUND OF THE INVENTION

[0003] Traditional techniques of studying the microscopic anatomy of a sample, in particular cellular or tissue structures, rely on methods in which the samples are stained in advance. A common staining method is, for example, Hematoxylin and Eosin (H&E) staining. Such staining methods usually require manual tasks as well as chemical and / or mechanical procedures on the sample, such as fixing, embedding in paraffin or wax, sectioning, and exposing to chemicals. These steps require lengthy sample preparation and need to be performed in a laboratory environment. The preparation of a Hematoxylin and Eosin (H&E) stain of the sample may take up to several days. The H&E-stain of the sample exhibits an enhanced contrast of certain cellular and / or tissue characteristics. For example, cell nuclei of the sample may be stained blue and the cytoplasm and extracellular matrix of the sample may be stained pink. Images showing those cellular and / or tissue characteristics are interpreted by an expert for assessing the sample. The expert may be, for example, but not limited to, a pathologist, a researcher, or a medical doctor. The assessment by the expert may, for example, be indicative of whether the sample contains cancerous tissue, i.e., a tumour.

[0004] For example, an imaging system proposed by Invenio Imaging employs Stimulated Raman Histology (SRH) for imaging a sample and displaying the results as a pseudo-colour-2- image. Multiple views are acquired and registered in an automated manner so that areas up to 10 x 10 mm2can be imaged (https: / / www.invenio-imaging.com / ).

[0005] A microscope may scan a sample in a point-by-point manner. At a scanning point, the sample is illuminated and one or more signals originating from the sample are detected. The one or more detected signals collected result in a pixel value for the scanning position. Repeating the illumination and detection for a several scanning positions results in an image of the sample or a portion thereof. A data acquisition time per scanning position depends on a dwell time of an illumination light beam at each scanning position. The image acquisition time (i.e., the sum of data acquisition times for the scanning positions of an entire image) depends on the total number of the scanning positions.

[0006] There is an interest in imaging larger samples. For example, such samples may have surface areas in the region of square centimetres. For a given resolution, the larger surface areas require the acquisition of data at an increased number of the scanning positions. The number of scanning positions scales quadratically with the surface area to be imaged, which results in the image acquisition time increasing in a quadratic manner.

[0007] As already mentioned, the acquired images are interpreted by the experts. The interpretation of the images usually comprises the identification of regions of interest (ROI), for instance, regions showing signs of cancer. The identification of one or more ROIs is usually performed through a visual assessment (“visual analysis”) by the expert, e.g., a pathologist. This step often requires a fair amount of time and high degree of expert knowledge. This step is moreover prone to human errors, e.g., due to exhaustion and overload of the expert.

[0008] Machine-learning approaches have shown potential use for image analysis and feature recognition in recent years. In particular, artificial neural networks have been successful in classifying various features within images. Trained machine-learning (ML) models based on artificial neural networks have also been useful in the field of pathology. ML models trained on H&E Whole Slide Images (WSIs) have, for example, shown promising results in accurately detecting cancerous regions. ML models, such as neural networks, have been trained mostly on high-resolution, traditionally stained images (H&E WSIs) showing individual nuclei and their size and shape.-3-

[0009] Using modem optical imaging techniques for acquiring images of the sample without the need of traditional staining promises an increased resource efficiency as well as improved diagnostic capabilities.

[0010] Several microscopic imaging techniques detect a weak photon flux resulting from light matter interactions, Stimulated Raman Scanttering (SRS), on the other hand, is a pumpprobe scheme that works in the high flux regime. SRS involves detecting a weak modulation signal (104- l 06) in a strong laser field, for which reason laser noise, e.g., quantum shot noise, has to be dealt.

[0011] For a given one of the dwell time of the illumination light beam at one scanning position, the modulation signal may be below the laser noise and thus cannot be recovered by post-processing averaging in the case of SRS, as opposed to, e.g., fluorescence microscopy.

[0012] The objective of the present invention relates to a method of imaging a sample that enables higher image acquisition speeds and enables increased accuracy of assessment of the imaged sample by experts or expert systems.

[0013] US 2020 / 341253 Al discloses a fluorescence microscope for stimulated emission depletion (STED) microscopy. In a method for imaging a sample, a confocal raw image is generated using excitation light and without using depletion light. In a subsequent step, a STED image is generated based on an analysis of the confocal raw image. The confocal raw image has a spatial resolution, which is determined by a spatial extent of the excitation light distribution with which the objective focuses into the sample, and which is lower than a resolution of the STED image.

[0014] US 2002 / 027203 Al discloses a confocal scanning microscope as well as method for examining a specimen. The method includes acquiring a preview image by scanning an illumination light beam across a specimen and marking a region of interest in the preview image. Subsequently, the marked region of interest is illuminated with an illumination wavelength selected by means of a spectrally selective element.

[0015] WO 2020 / 210746 Al discloses an imaging method, in which images are generated using two imaging modalities. The images are subsequently processed for analysis.

[0016] WO 2021 / 198243 Al discloses a method for virtually staining tissue samples, wherein a region of interest is identified by image processing.-4-

[0017] US 2010 / 225910 Al discloses a method for correlation spectroscopy, in which, based on an overview image, regions of interest of a sample are selected, which are subsequently examined using the correlation spectroscopy.

[0018] DE 10 2020 127320 B3 discloses a method for fluorescence microscopy, in which fluorescent molecules are localized in a sample by scanning. Prior to the scanning, scanning parameters are determined based on a raster image of the sample or a section of the sample.SUMMARY OF THE INVENTION

[0019] The present disclosure relates to a two-step assessment of at least one sample. The two-step assessment is based on a method of imaging the at least one sample. The two-step assessment (in the following also referred to as “assessment”) relates to a state of the at least one sample. The sample may be a biological sample, e.g., tissue or cell culture, or a non- biological sample.

[0020] In one aspect of the present disclosure, the state of the at least one sample relates to a presence of a disease in the at least one sample. The state of the at least one sample may be a diseased state or non-diseased state. The at least one sample may be assessed with respect to a presence of the diseased state.

[0021] The present disclosure addresses the need for faster and more efficient methods for imaging of at least one sample, including biological samples, in particular cellular or tissue structures, in the fields of research and medicine, e.g., for surgical operations. The method according to the present disclosure enables expediting the imaging process and, thus, reducing the need for extensive scanning processes.

[0022] An optical imaging system according to the disclosure enables improving a diagnosis based on the assessment of the at least one sample, based on the use of optical imaging techniques, pre-processing methods, and machine learning (ML) models. The optical imaging system according to the present disclosure reduces the sample preparation time.

[0023] The at least one sample may be obtained during a biopsy or a resection. The at least one sample may comprise cancerous tissue.

[0024] The present disclosure describes a method, an optical imaging system, a computer program (product), and a computer-readable (storage) medium for imaging at least on sample.-5-

[0025] According to the present disclosure, the method comprises several steps for imaging the at least one sample.

[0026] The imaging of the at least one sample comprises providing an illumination light beam. The illumination light beam comprises light of one or more illumination wavelengths for illuminating the at least one sample.

[0027] The method further comprises imaging the at least one sample at a plurality of first imaging positions. The imaging comprises illuminating, at the plurality of first illumination positions, the at least one sample with the illumination light beam; collecting first detection light from the plurality of first illumination positions; and generating first imaging data associated with the plurality of first imaging positions, wherein, for at least one of the plurality first imaging positions, the first imaging data is generated from several ones of the plurality of first illumination positions. The detection light may comprise light emitted, e.g., fluoresced, by the at least one sample, light transmitted by the at least one sample, light reflected by the at least one sample, and / or light scattered by the at least one sample.

[0028] For at least one of the plurality first imaging positions, the illumination light beam is moved across several ones of the plurality of first illumination positions during the collecting of the first detection light to generate the first imaging data, associated with the at least one of the plurality of imaging positions, from the several ones of the plurality of first illumination positions.

[0029] Any one of the plurality of first imaging positions may be associated with one or more associated ones of the plurality of first illumination positions.

[0030] The method further comprises analysing the first imaging data with respect to at least one detection wavelength, to identify at least one region of interest (RO I).

[0031] The method further comprises imaging the at least one ROI at a plurality of second imaging positions. The imaging of the at least one ROI at a plurality of second imaging positions comprises illuminating, at a plurality of second illumination positions, the at least one sample with the illumination light beam; collecting second detection light from the plurality of second illumination positions; and generating second imaging data associated with the plurality of second imaging positions. Any one of the plurality of second imaging positions may be associated with one or more associated ones of the plurality of second illumination positions.-6-

[0032] The method further comprises analysing the second imaging data with respect to the at least one detection wavelength.

[0033] The method according to the present disclosure enables the use of the first imaging data to gain an overview image of the at least one sample. The first imaging data may be generated at a first resolution, enabling fast acquisition of the first imaging data and fast gaining of the overview image. The overview image can provide a visual representation of the at least one sample. The overview image enables a first assessment of the at least one sample, e.g., identification of the at least one region of interest, e.g., in a clinical context.

[0034] In one aspect, the at least one sample is a tissue sample and the at least one ROI to be identified is a part of an organ or of a tumour that is of interest, e.g., to the pathologist. In another aspect, the at least one sample is a cell sample and the at least one ROI to be identified is a region where imaging conditions, e.g., relating to cell confluency, are suitable for conducting the method according to the disclosure. In yet a further aspect, the at least one sample an organoid (i.e., a 3D organ model) and the ROI to be identified is a z-plane (within the organoid) for more detailed imaging.

[0035] For example, the overview image may be converted into a virtual H&E overview image (also referred to as “virtually stained images”, “H&E-like images”, or “virtual H&E stain”), showing for example, but not limited to, cell bodies and cell nuclei and distributions thereof. The virtual H&E overview image enables an analysis, e.g., a visual analysis by a pathologist or an analysis based on image processing, without the need of prior manual H&E staining of the sample. This reduces the time required for a classification of the at least one sample, e.g., regarding the presence of diseased tissue, such as cancerous tissue, and thus accelerates the diagnosis.

[0036] The generation of the second imaging data enables gaining more details and information on the at least one ROI for further classification, e.g., more precise classification, of the at least one sample. The second imaging data enable gaining an in-depth image of the at least one ROI. The in-depth image has a higher resolution compared to the resolution of the overview image. The visual representation of the in-depth image may be converted into a virtual H&E in-depth image, showing for example, but not limited to, cell bodies and cell nuclei distributions which correspond to particular wavelengths in more detail than the overview image. The virtual H&E in-depth image based on the in-depth image enables a more detailed visual analysis, for example, by the pathologist or by means of image-7- processing, since further detail may be revealed due to the increased image resolution. The further detail enables improving the diagnosis. For example, the first imaging data may be missing some assessment-relevant detail, which the second imaging data may reveal. A second assessment of the at least one sample based on the second imaging data may detect the assessment-relevant detail, which improves the diagnosis in comparison with the first assessment based on the first imaging data.

[0037] The analysing of the first imaging data may comprise inputting the first imaging data into a trained machine-learning model.

[0038] The analysis of the first imaging data by means of a trained machine-learning model enables automatic detection of ROI of the at least one sample. The automatic detection enables assisting the pathologist, e.g., during surgery, resulting in the reduction of errors, e.g., human errors, such as false positives (e.g., wrongly assessing the presence of diseased tissue in the at least one sample) or false negatives (e.g., wrongly assessing the absence of diseased tissue in the at least one sample), and decreased diagnostic times.

[0039] The analysing of the first imaging data may comprise detecting at least one chemical marker. The at least one chemical marker may include an intrinsic marker, i.e., a chemical component of the at least one sample, which is detectable by means of at least one signature molecular vibration associated with the chemical component. The at least one chemical marker may further include externally introduced marker such as a fluorescent dye or a vibrational tag, also referred to as Raman-active marker. The detecting of the least one chemical marker enables assessing a presence in the at least one sample of the least one chemical marker.

[0040] The detecting of the at least one chemical marker may comprise detecting the at least one signature molecular vibration indicative of the at least one chemical marker. The spectrally detecting of a plurality of chemical markers further enables assessing relative amounts of the plurality of chemical markers.

[0041] In one aspect of the disclosure, the illumination light beam comprises light of two or more illumination wavelengths for exciting the at least one sample. The two or more illumination wavelengths may have one or more wavelength differences corresponding to one or more excitation energies of the at least one signature molecular vibration. The one or more wavelength difference are linked to one or more frequency differences of two or more optical frequencies corresponding to the two or more illumination wavelengths. These one-8- or more frequency differences also correspond to the one or more excitation energies of the at least one signature molecular vibration, which for a selected one of the at least one signature molecular vibration can be expressed as AE = hdv, where AE is the excitation energy of the selected one of the at least one signature molecular vibration, h is Planck’s constant, and Av is the frequency difference of the optical frequency v corresponding to the excitation energy of the selected one of the at least one signature molecular vibration. The illuminating with the at least two illumination wavelengths enable resonant excitation of the at least one chemical marker of the at least one sample. When the at least one signature molecular vibration is excited, an energy transfer from the illumination light beam to the at least sample corresponds to an amount of the excitation energy, which is transferred to the signature molecular vibration of the at least one chemical marker of the at least one sample.

[0042] The at least one detection wavelength may comprise the first illumination wavelength and / or the second illumination wavelength.

[0043] In one aspect, a first intensity of the detection light (e.g., the first detection light for gaining the overview image and / or the second detection light for generating the in-depth image) of the first illumination wavelength may be attenuated. A second intensity of the detection light (e.g., the first detection light and / or the second detection light) of the second illumination wavelength may be amplified (Raman amplification). By detecting the attenuation and / or the amplification, the signature molecular vibration may be detected (coherent Ramen scattering).

[0044] In a further aspect, the illumination light beam may comprise third illumination light of a third illumination wavelength. The third illumination light may be used to probe the resonant excitation by detecting a local increase in temperature (Stimulated Raman Photothermal effect).

[0045] The analysing of the second imaging data may comprise inputting the second imaging data into a / the trained machine-learning model.

[0046] The analysing of the second imaging data by means of the trained machine-learning model enables automatically classifying the at least one sample, e.g., regarding the presence in the at least one sample of diseased tissue. For example, the classification of the at least one sample resulting in a false-positive assessment (e.g., wrongly diagnosing the presence of the diseased tissue) is reduced by the analysis by means of the trained ML model. Furthermore, the second imaging data having a higher resolution than the first imaging data,-9- i.e., revealing the further detail of the at least one sample enables a further assessment of the at least one ROI. The further assessment of the at least one ROI reduces the likelihood of wrongly classifying the at least one sample.

[0047] The plurality of first imaging positions is associated with a first nearest-neighbour distance. The plurality of second imaging positions is associated with a second nearest- neighbour distance. The first nearest-neighbour distance of the plurality of first imaging positions may be larger than the second nearest-neighbour distance of the plurality of second imaging positions. The first nearest-neighbour distance having a larger value than second nearest-neighbour distance, enables reaching a rapid first classification of the at least one sample, i.e., identifying the at least one ROI. In particular, a large one of the at least one sample may thus be classified, regarding the identification of the at least one ROI, in the first assessment (or first assessment step) in a fast manner. The second nearest-neighbour distance having a smaller value that the first nearest-neighbour distance, enables a more detailed analysis of the at least one ROI, in the second assessment (or second assessment step). The two-step assessment, based on the first nearest-neighbour distance being larger than the second nearest-neighbour distance, enables reducing diagnostic times.

[0048] The first nearest-neighbour distance may be larger than 5 pm. The second nearest- neighbour distance may be equal to or smaller than 5 pm.

[0049] The method may further comprise user-based identifying, based on the first imaging data, of at least one further ROI.

[0050] The user-based identifying of the at least one further ROI enables user-based interaction. For example, the pathologist or researcher may identify, and manually select, a particular region of the at least one sample to indicate the further ROI. The user-based identification may be based on the already obtained first imaging data or may be independent from the obtained first imaging data. The process of selecting the at least one further ROI may be done by marking the further ROI on the overview image. The user-based identifying of at least one further ROI improves the usability and flexibility of the method while imaging the sample and enables synergies of the machine-learning-based assessment and the pathologist’s experience.

[0051] The acquisition of the further second imaging data for the further ROI identified by the user enables obtaining yet further detail of the at least one sample. This helps to improve the classification of the at least one sample as well as diagnostic decisions.-10-

[0052] The method may further comprise storing the first imaging data generated for the plurality of first imaging positions as first image data, and / or the second imaging data generated for the plurality of second imaging positions as second image data.

[0053] The storage of the first imaging data and the second imaging data enables usage for subsequent image processing, training procedures, automatic feature recognition, visualizations, and documentation.

[0054] The method may further comprise displaying the first image data and / or the second image data.

[0055] The displaying of the first image data, for example the overview image, on a screen in an operating room or remotely enables the intervention of the pathologist in the classification of the at least one sample. The first image data may be processed and converted into the virtual H&E image. The display of the first image data may allow the pathologist to identify the at least one ROI and to make diagnostic decisions.

[0056] The method may further comprise visually analysing the first image data and / or the second image data. The visually analysing may be done by means of a graphical userinterface. The graphical user-interface may provide, for example, interactive control elements for magnifying portions of the overview image. The generating of the first imaging data may comprise integrating at least one electrical signal generated from the first detection light.

[0057] An integration time of the integrating of the at least one electrical signal may be adapted to a scanning frequency / scan of the moving of the illumination light beam across the several ones of the plurality of first illumination positions.

[0058] The integration time of the integrating of the at least one electrical signal may be set such that a predefined signal -to-noise ratio is achieved.

[0059] The method may further comprise moving the at least one sample.

[0060] The present disclosure also describes an optical imaging system for imaging at least one sample. The optical imaging system for imaging the at least on sample comprises an illumination module, a control unit, a detection module, a memory, and at least one processing unit. The illumination module is configured to provide an illumination light beam for illuminating the at least one sample at at least one illumination position. The illumination light beam comprises light of one or more illumination wavelengths. The control unit is configured to move the illumination light beam to the at least one illumination position. The-11- detection module is configured to collect the detection light from the at least one illumination position (the first detection light and / or the second detection light) and to convert the detection light into imaging data. The detection module may detect the detection light by means of a detection objective lens. The detection module may detect the detection light in a back-scattering direction (i.e., opposite an illumination direction in which the illumination light beam illuminates the at least one sample). The memory is configured to store the imaging data received from the detection module as image data. The processing unit is configured to analyse the imaging data with respect to at least one detection wavelength to identify at least one region of interest (ROI). The control unit is configured, for at least one of the plurality of imaging positions, to move, during collecting of the detection light by the detection module, the illumination light beam across several ones of the plurality of illumination positions to generate the data, associated with the at least one of the plurality of imaging positions, from the several ones of the plurality of illumination positions.

[0061] The optical imaging system enables fast automated imaging of the at least one sample to visualize the at least one sample, such as, but not limited to, cellular or tissue structures of the at least one sample with respect to the interaction of the illumination light beam with the sample indicative of the presence of a chemical marker. The system may be beneficial in surgeries, for example, in lumpectomy surgery.

[0062] The detection module comprises at least one detector. The at least one detector has a detector bandwidthdet.

[0063] According to one aspect of the disclosure, the detection module may comprise one or more optical filters to selectively filter the detection light (the first detection light and / or the second detection light). The one or more filters may be arranged such that the detection light from the at least one sample reaches the one or more filters before being collected by at least one detector (see below). The one or more optical filters may be configured to filter the detection light with respect to the at least one detection wavelength and / or one or more ranges of the least one detection wavelength. In one aspect, the one or more optical filters may be configured to filter the detection light with respect to the one or more illumination wavelengths.

[0064] According to one aspect of the disclosure, the detection module may further comprise at least one spectrometer unit. The at least one spectrometer unit may comprise a grating and / or a prism for partitioning the detection light (e.g., the first detection light and / or the-12- second detection light) dependent on the at least one detection wavelength. The detection light from the at least one sample may be partitioned into one or more components, the one or more components each containing light of selected ones of the at least one detection wavelength. In one aspect, the one or more components may each comprise a selected one of the one or more illumination wavelengths. In a further aspect, the one or more components may each be directed to several ones of the at least one detector. The at least one spectrometer unit enables detecting a plurality of signature wavelengths indicative of a plurality of chemical markers.

[0065] The processing unit may be configured to detect at least one detection frequency f of at least one electrical signal (photocurrent) generated by the at least one detector indicative of at least one chemical marker.

[0066] The at least one detection frequency f may be associated with a frequency of modulation of the detection light. For example, the detection light, the detection wavelength of which is equal to the first illumination wavelength or the second illumination wavelength, may be modulated due to the interaction of the illumination light beam with the at least one sample. This modulation of the detection light may have a modulation frequency. In one aspect of the disclosure, the modulation frequency f may for example lie in a range from 100 kHz and 40 MHz.

[0067] In one aspect, a balanced detection approach may be implemented. In this approach, the processing unit subtracts a noise background from the at least one electrical signal (photocurrent) generated by the at least one detector. The subtracting of the noise background is also referred to as denoising. The noise background may be determined by directing the illumination light beam directly to the at least one detector (referred to as reference beam) and detecting a generated reference electrical signal (reference photocurrent). The reference beam does not undergo any interaction with the at least one sample.

[0068] The optical imaging system may further be configured to input the image data into at least one trained machine-learning model.

[0069] The trained machine-learning model enables automatic detection of the least one ROI within the sample. The at least one ROI may be, for instance, but not limited to, a region of the at least one sample containing diseased, e.g., cancerous, tissue. The automatic detection-13- provides assistance to the pathologist during surgery and enables reducing errors and diagnostic times.

[0070] The optical imaging system may further comprise at least one display device. The at least one display device, communicatively coupled with the at least one processing unit, is configured to display an image based on the image data. The at least one ROI may be indicated in the image.

[0071] The optical imaging system may further be configured to provide a graphical user interface for visually analysing the image data by means of the at least one display device.

[0072] The displaying of the image on the at least one display device enables visually classifying by the pathologist of the at least one sample, e.g., in an operating room where the surgery takes place, or remotely.

[0073] The present disclosure further relates to a computer program product. The computer program product comprises instructions, which cause, when the program is executed by a processing unit of an optical imaging system, the optical imaging system to perform a method according to the present disclosure.

[0074] The computer program product enables the automation of the method for imaging the at least one sample and reduces the image acquisition time.

[0075] The present disclosure further describes a computer-readable storage medium.

[0076] The computer-readable storage medium has stored thereon the computer program product.DESCRIPTION OF THE FIGURES

[0077] FIG. 1 describes a flow chart of an imaging method for imaging a sample.

[0078] FIG. 2 discloses a table indicating wavenumbers corresponding to some cancer markers.

[0079] FIG. 3 illustrates illumination positions for illuminating the at least one sample and imaging positions at which first imaging data of a sample is acquired.

[0080] FIG. 4 describes a flow chart of a method for training an ML model based on virtual H&E images for classifying at least one sample.

[0081] FIG. 5 A shows a H&E WSI of a sample with a magnified excerpt.

[0082] FIG. 5B illustrates a sub-sampling routine for reducing the resolution of an image.-14-

[0083] FIG. 5C illustrates the sub-sampling routine applied to a portion of the sample.

[0084] FIG. 6 shows two overview images based on pre-processed H&E WSIs, the left overview image containing the ground truth and the right overview image having been analysed by a trained ML model.

[0085] FIG. 7 illustrates a use case for the presented imaging method and optical imaging system.

[0086] FIG. 8 shows a schematic illustration of the optical imaging system.

[0087] FIG. 9 shows first imaging data represented as a virtual H&E image with a first nearest-neighbour distance of 20 pm.

[0088] FIG. 10 shows an example of a scanning movement.DETAILED DESCRIPTION OF THE INVENTION

[0089] The invention will now be described on the basis of the figures. It will be under-stood that the embodiments and aspects of the invention described herein are only examples and do not limit the protective scope of the claims in any way. The invention is defined by the claims and their equivalents. It will be understood that features of one aspect or embodiment of the invention can be combined with a feature of a different aspect or aspects and / or embodiments of the invention.

[0090] FIG. 1 describes a flow chart of a method S for imaging at least one sample 500. The method S comprises providing S100 an illumination light beam 812, imaging SI 10 the at least one sample 500, analysing S120 the first imaging data and identifying at least one region of interest (ROI), imaging S130 the at least one ROI of the at least one sample 500, and analysing S140 the second imaging data.

[0091] The provision SI 00 of illumination light is implemented by an optical imaging system 710 for imaging the at least one sample 500. The optical imaging system 710 comprises an illumination module 712, a scanner 813, an illumination objective lens 814, a detection module 830, a control unit 820, a memory 840, and a processing unit 713.

[0092] The illumination module 712 is configured to provide the illumination light beam 812 for illuminating the at least one sample 500 at at least one illumination position 320 (see FIG. 3).-15-

[0093] In a further aspect, the optical imaging system 710 further comprises a stage 550 for moving the sample 500 (see FIG. 8). In this aspect, the sample 500 is supported by the stage 550. The stage 550 is configured to be movable. The stage 550 may be controlled by the control unit 820 to be moved.

[0094] The movability of the stage 550 enables image mosaicking (also termed image tiling) for generating an enlarged image of the sample 500. This enlarged image may have an enlarged field of view (enlarged FOV). The enlarged image of the sample may be composed of several ones of the first image described below or composed of several ones of the second image described below. The image mosaicking may comprise stripe-based mosaicking.

[0095] The enlarged image may be constructed by aligning and / or juxtaposing several ones of the first image or several ones of the second image. Ones of the first image or of the second image, for which the corresponding fields of view (FOV) overlap, may be aligned and / or juxtaposed. Image mosaicking enables improved visualization of an anatomical site for biomedical imaging. Image mosaicking may for example assist a surgeon during invasive surgery.

[0096] The illumination of the at least one sample 500 may result in an interaction of the illumination light beam with the at least one sample 500. Examples of the interaction are an excitation of a signature molecular vibration associated with a chemical component of the at least one sample 500 or an excitation of a fhiorophore.

[0097] The illumination module 712 comprises a light source 810 for generating the illumination light beam 812. The light source 810 comprises at least one laser device. An example of the at least one laser device is continuous-wave laser device or a pulsed-laser device. It will be understood that the light source 810 is not limited to the laser device may comprise any suitable device for generating light for the illumination light beam 812.

[0098] In one aspect, the light source 810 may comprise a first laser device and a second laser device. The light source 810 comprising the first laser device and the second laser may be dual -frequency laser device, such as disclosed in US 8 681 331 B2, the disclosure of which is hereby incorporated herein by reference. The first laser device may provide first illumination light at a first illumination wavelength (also referred to as pump wavelength or Stokes wavelength). The first illumination light has a first spectral bandwidth. The first spectral bandwidth may be, for example, 2 nm or less (“narrow bandwidth”), in which case the first laser device is a picosecond laser. The second laser device may further provide-16- second illumination light at a second illumination wavelength different from the first wavelength. The second illumination light may have a second spectral bandwidth. In one example, the second spectral bandwidth may be 2 nm or less (“narrow bandwidth”).. In another example the second spectral bandwidth may be broader than 2 nm or less, in which case the second illumination light provided by the second laser (“broadband laser”) may be spectrally spread by a grating and measured by the at least one detector 834 of the detection module 830.

[0099] In another aspect, the light source 810 may further comprise a third laser device providing third illumination light at a third illumination wavelength. The third illumination wavelength may be different from the first illumination wavelength and second illumination wavelength. The third illumination wavelength may be in the visible part of the spectrum to achieve a higher resolution. The third illumination light may have a third spectral bandwidth that different from, or equal to, one or both of the first spectral bandwidth and the second spectral bandwidth.

[0100] The detection module 830 may comprise at least one detector 834 for receiving detection light 816. The detection module 830 may comprise a spectrometer unit 832. The at least one detector 834 comprises at least one pixel detector 834. Examples of the at least one pixel detector 834 are a photodiode or a pixel array. The pixel detector 834 may detect the detection light 816 for a plurality of imaging positions 330 (see below) in a sequential manner. The at least one pixel detector 834 has a detector bandwidth d / LIA.

[0101] In one aspect of the disclosure, the processing unit 713 further comprises a lock-in amplifier 716 (see FIG. 8). The lock-in amplifier 716 may receive an electrical signal (or photocurrent) from the at least one detector 834. Alternatively, the lock-in amplifier 716 may receive the electrical signal (or photocurrent) from the memory 840 as store the data received from the detection module 830

[0102] The lock-in amplifier 716 may have a lock-in amplifier bandwidth dLIA. The lock-in amplifier bandwidth d / LIAof the lock-in amplifier 716 is associated with an integration time of the lock-in amplifier 716: T = 1 / dLIA. The lock-in amplifier bandwidth LIA °f the lock-in amplifier 716 is smaller than the detector bandwidth d / LIAof the at least one pixel detector 834: d / LIA< ^ / det-

[0103] The integration time r of the lock-in amplifier 716 may be set such that demodulating of a the modulation transfer of the electrical signal (photocurrent) of the at-17- least one detector 834 and / or rejection of the shot noise is enabled. The integration time r of the lock-in amplifier 716 is inversely proportion to the signal -to-noise ratio (SNR) of the electrical signal (photocurrent). For example, the relationship of the integration time r and the SNR may be described by the formula: SNR = ?2 / 8 T RIN / / / / avg), where is a relative intensity gain of the Stokes beam, / avgis an average electrical signal (average photocurrent), and RIN / / 0, / avg) is a relative intensity noise of the electrical signal (photocurrent) as a function of the frequency f and / avg(see, e.g., DOI: 10.1063 / 1.5129212). The setting of the integration time r of the lock-in amplifier 716 in this way enables extracting one or more of a phase difference, an amplitude, and / or a product of the phase and the amplitude (see below).

[0104] In one aspect of the disclosure, the integration time r may be set in relation to a scanning frequency / scan (or movement frequency) of the moving (scanning) of the illumination light beam 812 across several ones of the plurality of first illumination positions 320-1. The the scanner 813 may be controlled, e.g., by the control unit 820, to move the illumination light beam 812 at the scanning frequency can. The integration time r may be set based on the following formula: T = c • N • r / (LFOV ■scan). This formula provides a relationship between the integration time r and the scanning frequency can, where c is an implementation-dependent pre-factor, TV is the number of first illumination positions 320-1, LFOV is a linear extension of the field of view along a scanning path of the illumination light beam 812, and r is a spot size or focus size (see below) of the illumination light beam 812. Examples of values for the afore-mentioned variables are: r = 310nm, / can= 100Hz, LFOV = 500pm. A size (or dimension) of the corresponding imaging position 330 is N ■ r.

[0105] The lock-in amplifier 716 may output a product of an amplitude of the electrical signal (or photocurrent) at a reference frequency of a reference electrical signal. The lock-in amplifier 716 may further output phase difference between the electrical signal and reference electrical signal.

[0106] In one aspect, the lock-in amplifier 716 further comprises a two-phase lock- in amplifier. This two-phase lock-in amplifier may output a so-called in-phase component and a quadrature component, from which both the amplitude of the electrical signal at the reference frequency of the reference electrical signal and phase difference between the electrical signal and reference electrical signal may be calculated.-18-

[0107] The lock-in amplifier enables phase-sensitive detection. Moreover, the lock- in amplifier 716 enables detecting a weak one of the electrical signal buried in noise, e.g., shot noise, in particular quantum shot noise. For example, the lock-in amplifier 716 may enable detecting the electrical signal at a SNR of -60 dB or less.

[0108] The electrical signals may typically be rather low, for example 10'5, thus calling for modulation-transfer techniques: one beam is modulated in amplitude, and the signal is detected on the other beam via a lock-in amplifier.

[0109] In one aspect of the disclosure, the detection module 830 may be configured to receive detection light in a back-scattering direction through the illumination objective lens 814 (such as is disclosed in US 8 681 331 B2).

[0110] In another aspect, the optical imaging system 710 may comprise a detection objective lens 815. The detection objective lens 815 may be arranged around or in the vicinity of the illumination objective lens 814.

[0111] The memory 840 is configured to store the data received from the detection module 830. The processing unit 713 is configured to analyse the data.

[0112] The light of the illumination light beam 812 is directed onto the at least one sample 500 so that the at least one sample 500 is illuminated at the at least one illumination position 320. At the at least one illumination position 320, the illumination beam 816 has a spot size (or focus size). The spot size may be defined by a diameter of a point spread function of the illumination light beam 812, e.g., at a focus of the illumination beam 816. The spot size or focus size may be kept constant during an imaging.

[0113] In one aspect of the invention, the spot size or focus size, e.g., an extent of a point spread function of the illumination light beam 812 focussed on the at least one sample 500, is approximately 500 nm. It will be understood that this spot size or focus size is a mere example and that other spot sizes different to 500 nm are also applicable.

[0114] The illumination light of the provided illumination light beam 812 comprises light having the one or more illumination wavelengths. The one or more illumination wavelengths are preferably in an illumination wavelength range of 300 nm to 1700 nm. It will be understood that other wavelengths outside the illumination wavelength range of 300 nm to 1700 nm may be used.

[0115] In another aspect, the one or more illumination wavelengths of the illumination light beam 812 are selected with respect to a chosen optical imaging technique.-19-In Stimulated Raman Scattering (SRS) microscopy, for instance, two illumination wavelengths are provided to illuminate the at least one sample 500 at the at least one illumination position 320. The two illumination wavelengths are referred to as “pump wavelength” and “Stokes wavelength”. The pump wavelength and the Stokes wavelength may have a chosen wavelength difference. The chosen wavelength difference corresponds to an energy difference of a photon of the pump wavelength and a photon of the Stokes wavelength. In one aspect, the pump wavelength may be in a pump wavelength range of 700 nm to 980 nm, and the Stokes wavelength may be in a Stokes wavelength range of 1020 nm to 1060 nm. The pump wavelength range and the Stokes wavelength range are subranges of the illumination wavelength range. At least one of the two illumination wavelengths may be adjustable with respect to its illumination wavelength within the preferable range. The aforementioned range of wavelengths approximately corresponds to a range of wavenumbers of 400 cm’1to 4500 cm'1. It will be understood that other wavelengths of the pump wavelength and the Stokes wavelength are also applicable. For example, the pump wavelength and / or the Stokes wavelength may lie in the wavelength range 300 nm to 1700 nm, as mentioned above.

[0116] The energy difference between the pump wavelength and the Stokes wavelength may be selected based on a selected vibrational or rotational transition (energy transition) of a molecule from a first state, e.g., a ground state, to a second state, e.g., a vibrational (or rotational) state, different from the first state. The at least one sample 500 may contain a plurality of the molecule.

[0117] The energy transition of the molecule from the ground state to the vibrational (or rotational) state may be resonantly enhanced when the chosen wavelength difference (corresponding to the energy transition and termed signature wavelength) substantially equals an energy absorbed or emitted by the molecule during the energy transition. The resonant enhancement may result in a signal (e.g., light emitted or absorbed by the at least one sample 500) associated with the signature wavelength that is detectable.

[0118] For example, in the event of resonant enhancement, an intensity of the light of the pump wavelength (a component of the illumination light beam 812, as explained above) may have reduced after passing through the at least one sample 500, due to enhanced absorption of the light of pump wavelength by the plurality of the molecule.-20-

[0119] Additionally or alternatively, an intensity of the light of the Stokes wavelength (a further component of the illumination light beam 812, as explained above) may have increased after passing through the at least one sample 500, due to enhanced emission / transmission of the light of Stokes wavelength by the plurality of the molecule present in the at least one sample 500. The detectable signal is indicative of a presence of the molecule in the at least one sample 500. It is thus possible to derive from the detectable signal the presence of the molecule in the at least one sample 500. The molecule may be a chemical marker indicative of the presence of diseased tissue in the at least one sample 500.

[0120] It may further be possible to derive from the detectable signal a degree of the presence of the molecule in the at least one sample 500. For example, an amplitude of the detectable signal (which corresponds to the afore-mentioned intensity of the light of the pump wavelength) may indicate an amount of the molecule present in the at least one sample 500. The amount may be a relative amount. The relative amount may indicate that the molecule is present in the at least one sample 500 to a higher, equal, or lesser degree than another one of the molecule.

[0121] For example, SRS microscopy enables detecting the presence of the molecule, e.g., the chemical marker, based on a signature wavelength in a detection wavenumber range from 300 cmA-l to 3100 cmA-l. The detection wavenumber range comprises a first detection wavenumber subrange from 300 cmA-l to 1750 cmA-l (referred to as “fingerprint region”) and a second detection wavenumber subrange from 2750 cmA-l to 3100A-1 (referred to as the “C-H stretch region”). In one aspect of the disclosure, at least one first chemical marker may be detected in the first wavenumber subrange. Additionally or alternatively, at least one second chemical marker may be detected in the second wavenumber subrange. Detecting one or more chemical markers allows to increase the quality and scope of the assessment (e.g., a diagnostic assessment) of the at least one sample 500, since more details and information are available regarding the at least one sample 500 and a disease potentially present in the at least one sample 500. Other detectable signals with wavenumbers outside the aforementioned detection wavenumber range (or its subranges) may also be detected and used to obtain information regarding the at least one sample 500. The SRS therefore enables inferring the presence of the chemical markers in the at least one sample 500.-21-

[0122] Investigating the at least one sample 500 with respect to the presence of one or more molecules may be used to examine the at least one sample 500 for so-called cancer markers (examples of the molecule or chemical marker). The cancer markers present in cancerous tissue may be detected based on the presence of the chemical bonds in CH2 groups and CH3 groups. The CH2 groups and the CH3 groups are associated with the presence in the at least one sample 500 of lipids and proteins, respectively. A distribution and / or density of the lipids and / or the proteins can in turn be associated with cell bodies and cell nuclei, respectively.

[0123] SRS techniques may thus be used to highlight the distribution and / or the density of cell bodies and cell nuclei in images of the at least one sample 500, e.g., for examination by experts. For example, cell mitosis (i.e., the generation of a new cell nucleus for cell division) and / or aspects and sizes of cell nuclei are helpful in the assessment of the at least one sample 500, which comprises potentially cancerous tissue, and facilitate diagnosis. The distinction between cell bodies and cell nuclei can be achieved by employing an SRS device and collecting the detection light 816 from the at least one sample 500 at two wavenumbers: 2845 cmA-l and 2930cmA-l corresponding to the CH2 and the CH3 chemical bonds, respectively.

[0124] The research article “Stimulated Raman histology: one to one comparison with standard hematoxylin and eosin staining" (Sarri et al., Vol. 10, No. 10 / 1 October 2019 / Biomedical Optics Express, 5378- 5384) shows images of a tissue sample comprising cell bodies and cell nuclei as described above. The research article also shows a comparison between an image of an H&E stain of the tissue sample and an image of the same tissue sample acquired by a combination of SRS and second harmonic generation (SHG). A setup up with two pump beams and one Stokes beam was used. The two pump beams were operated at the wavelengths of 797.3 nm and 792.2 nm which correspond to vibrations of chemical bonds in CH2 groups (detected at a wavenumber of 2845 cmA-l) and chemical bonds in CH3 groups (detected at wavenumber 2930 cmA-l), respectively. The Stokes beam was operated at 1031 nm.

[0125] It will be understood that other cancer markers are also known, for example, but not limited to, those ones listed in the table of FIG. 2 (see Cicerone in 2018, “Histological coherent Raman imaging: a prognostic review.” Analyst 143, 33-59). It will also be understood that signals of other chemical markers may also be detected and may serve as-22- indicators for different characteristics of the at least one sample 500, i.e., characteristics which are not associated with cancerous tissue. Generally, the chemical markers may be chosen dependent on a disease, for which the at least one sample 500 is to be classified.

[0126] In one aspect of the disclosure, an SRS (stimulated Raman scattering) microscope is employed. The SRS microscope may comprise one or more light sources 810 providing more than one pump wavelength and / or more than one Stokes wavelength for providing the illumination light beam 812 capable of illumination and / or exciting the at least one sample 500 with respect to more than one chemical marker (e.g., more than one cancer marker). This enables the simultaneous detection of two or more chemical markers. In other words, instead of a sequential data acquisition, i.e., acquiring data of one chemical marker at a time, this aspect of the disclosure enables parallel detection of several chemical markers. This improves a data acquisition time and an image acquisition time (i.e., the sum of data acquisition times for imaging positions 330).

[0127] The several chemical markers may have associated wavenumbers (or wavelengths) within the fingerprint region and within the C-H stretch region.

[0128] In another aspect of the invention, during data acquisition, the pump beam and / or the Stokes beam of the illumination light beam 812 is / are continuously tuned so that signals from the entire fingerprint region and from the range of 2750 cmA-l to 3100 cmA-l may be detected at the at least one illumination position 320.

[0129] Other optical imaging techniques may also be applicable to provide an illumination light beam 812 for illuminating and / or exciting the at least one sample 500. The other optical imaging techniques may comprise, for example, but not limited to, confocal reflectance microscopy, multiplexed stain-free microscopy (MUSE), photoacoustic imaging, multiphoton imaging, endoscopic in-vivo fluorescence imaging, radio frequency spectroscopy, fluorescence-based confocal laser scanning microscopy, light sheet microscopy, mid-infrared spectroscopy, and non-linear microscopy. It will also be understood that, for optical imaging techniques other than SRS microscopy, different parameters may be set and / or adjusted, e.g., the wavelengths of the light of the illumination light beam 812, for detection of the chemical markers.

[0130] The method S according to the disclosure comprises imaging SI 10 (see FIG. 1) the at least one sample 500. The imaging of the at least one sample 500 comprises moving (scanning) the illumination light beam 812 from one scanning point to another scanning-23- point, i.e., from one of the plurality of illumination positions 320 to another one of the plurality of illumination positions 320. The moving (scanning) of the illumination light beam812 may be performed by the scanner 813.

[0131] In one aspect of the disclosure, the scanner 813 comprises a first galvanometric mirror (not shown) for moving (scanning) the illumination light beam 812 along a first scanning axis (e.g., one of the vertical direction DI and the horizontal direction D2 described below with reference to FIG. 5B). In this aspect, the scanner 813 further comprises a second galvanometric mirror (not shown) for moving (scanning) the illumination light beam 812 along a second scanning axis (e.g., the other of the vertical direction DI and the horizontal direction D2).

[0132] The illuminating SI 10 comprises illuminating the at least one sample 500 at a plurality of first illumination positions 320-1 (see FIG. 3). At each of the first illumination positions 320-1, the illumination light beam 812 illuminates the at least one sample 500 such that the one or more molecules (chemical markers) of interest are illuminated. The illuminating of the at least one sample 500 may be performed by means of the illumination objective lens 814, which projects light received from the light source 810 onto the at least one sample 500. The illuminating of the at least one sample at the plurality of first illumination positions 320-1 may comprise the moving (scanning) of the illumination light beam 812 to the plurality of first illumination positions 320-1. The scanner 813 is configured to position the illumination light beam 812 such that the at least one sample 500 is illuminated at the plurality of first illumination positions 320-1. In one aspect, the scanner813 may position the illumination light beam 812 at a single one of the plurality of first illumination positions 320-1 at a time. In another aspect, the scanner 813 moves or sweeps the illumination light beam 812 across several ones of the plurality of first illumination positions 320-1 at a time.

[0133] The plurality of first illumination positions 320-1 may be represented by a set of pairs of an x-coordinate and a y-coordinate, defined with reference to an x-y-coordinate system. The plurality of first illumination positions 320-1 may be predefined, e.g., by a default setting or based on prior knowledge of the at least one sample 500.

[0134] A distance of two adjacent ones of a plurality first imaging positions 330-1 may be dependent on the spot size (e.g., 500 nm) of the illumination light beam 812. In another aspect, the spot size may be dependent on at least one dimension of the at least one-24- sample 500 (sample size) and / or the data acquisition time or a dwell time for generating the first imaging data and / or the second imaging data. The data acquisition time and / or the dwell time for generating the first imaging data and / or the second imaging data may depend on the at least one dimension of the at least one sample 500. Alternatively or additionally, the data acquisition time for generating the first imaging data and / or the second imaging data may depend on an nearest-neighbour distance.

[0135] For any one of the plurality of first imaging positions 330-1, a nearest- neighbour distance may be defined by a distance to at least one nearest neighbouring one of the plurality of first imaging positions 330-1. Alternatively, the nearest-neighbour distance may be defined as an average distance to several neighbouring ones of the plurality of first imaging positions 330-1. A coarseness (or fineness) of a grid, formed by the plurality of first imaging positions 330-1, may be defined, for example, based on an average or a minimum of the nearest-neighbour distance across the plurality of first imaging positions 330-1.

[0136] In another aspect of the invention, a first nearest-neighbour distance of the plurality of first imaging positions 330-1 may be greater than 5 pm. In another aspect of the invention, the first nearest-neighbour distance is 10 pm. In another aspect of the invention, the first nearest-neighbour distance is 20 pm.

[0137] In one aspect, the plurality of first imaging positions 330-1 may be defined or pre-defined such that the coarseness (or fineness) of the grid formed by the plurality of first imaging positions 330-1 determines a data acquisition time. Reducing the data acquisition time enables accelerating a surgery, e.g., a breast cancer surgery.

[0138] In another aspect, the plurality of first imaging positions 330-1 may be predefined such that the coarseness (or fineness) of the grid is determined by the plurality of first imaging positions 330-1 such that a risk of false negatives (i.e., wrongly classifying the at least one sample 500 as not containing diseased tissue, e.g., cancerous tissue, which would result in wrongly assessing an absence of the at least one ROI, e.g., due to insufficient information) is reduced.

[0139] The imaging SI 10 comprises collecting the detection light 816 from the at least one sample 500, e.g., from the one or more illuminated ones of the plurality of first illumination positions 320-1. The collecting of the detection light 816 may be performed by the detection objective lens 815.-25-

[0140] In one aspect of the disclosure, the illumination objective lens 814 and the detection objective lens 815 may be one and the same objective lens. In another aspect, the illumination objective lens 814 and the detection objective lens 815 may be arranged on opposite sides with respect to the at least one sample 500 (i.e., in a 180-degrees arrangement). In a further aspect, the illumination objective lens 814 and the detection objective lens 815 may be arranged perpendicular with respect to one another. The illumination objective lens 814 and the detection objective lens 815 may be arranged perpendicularly with respect to the at least one sample 500. The illumination objective lens 814 and the detection objective lens 815 may form a single objective 814, 815 lens for illuminating the at least one sample 500 and for detecting the detection light 816 from the at least one sample 500. The single objective lens 814, 815 enables application of the method S according to the disclosure in back-scattering direction.

[0141] The imaging SI 10 comprises generating first imaging data, associated with the corresponding one of the plurality of first imaging positions 330-1, based on the collected light. The collected light may be spectrally resolved. The collected light may be associated with detected ones of the one or more molecules (chemical markers). The generating of the first imaging data may be performed by the at least one pixel detector 834 of the detection module 830. The at least one pixel detector 834 may be part of the spectrometer unit 832. The spectrometer unit 832 may be an imaging spectrometer.

[0142] The collection of the detection light 816 from the at least one sample 500 takes a detection time. The detection time, e.g., for detecting the signal at one of the first imaging positions 330-1, e.g., to obtain a sufficiently strong signal or to read out the at least one detector 834, multiplied by the number of first imaging positions 330-1, results in the data acquisition time. The detection time may be in the microsecond range, for example, but not limited to, 10 ps or 40 ps.

[0143] For any one of the plurality of first imaging positions 330-1, the first imaging data may comprise first imaging data elements associated with any one of the one or more wavenumbers (that are associated with the chemical markers). The first imaging data elements may comprise intensities of the collected light for the corresponding wavenumber. The generated first imaging data at any one of the plurality of first imaging positions 330-1 may be used to generate one or more values. Based on the one or more values and the corresponding one of the plurality of first imaging positions 330-1, a first image of the at-26- least one sample 500 may be constructed. The first image may be the overview image of the at least one sample 500.

[0111] FIG. 3 schematically shows an example of proportions of dimensions (or sizes) of the first imaging positions 330-1 (represented by bold squares) in relation to the spot size (or focus size, e.g., defined by a point spread function) of the illumination light beam 812 at the at least one sample 500. The focus size of the illumination light beam 812 may have dimensions LI and L2 that are smaller than the dimensions of the first imaging positions 330-1. In the non-limiting example shown in FIG. 3, the imaging positions 330-1 each correspond to 16 ones of the plurality of first illumination positions 320-1 which have an individual size corresponding to the focus size of the illumination light beam 812. In this non-limiting example, only four of the 16 ones of the plurality of first illumination positions 320-1 contribute to the detection light 816 for the corresponding one of the plurality of first imaging positions 330-1. The plurality of first illumination positions 320-1 are shown here to be non-overlapping. In another aspect, the plurality of first illumination positions 320-1 may be overlapping. In other words, the focus size may be smaller than the first nearest- neighbour distance of the first imaging positions 330-1. The first nearest-neighbour distance of the first imaging positions 330-1 may be 20 pm or 10 pm. The first imaging positions 330-1 may correspond to clusters of illumination positions 320, wherein the illumination positions 320 of the clusters may be used for illuminating the at least one sample 500.

[0145] In an aspect of the disclosure, the illumination light beam 812 is moved (scanned) for the data acquisition of the first imaging data for a selected one of the plurality of first imaging positions 330-1. The illumination light beam 812 is moved during the collection of the detection light 816. The illumination light beam 812 may be moved such that more than one of the illumination positions 320 (several ones of the illumination positions 320 in the cluster which corresponds to the selected imaging position 330) contribute to the detection light 816. For example, the illumination light beam 812 may be moved along a horizontal line 331, a vertical line 332, or a diagonal line 333 across at least a portion of the cluster of illumination positions 320 associated with the selected one of the plurality of first imaging positions 330-1. More generally, the illumination light beam 812 may be moved along an arbitrary path across at least the portion of the cluster of illumination positions 320. The detection light 816 is collected for those illumination positions 320 across which the illumination light beam 812 passes during the movement of the illumination light-27- beam 812. Ather the collection, the detection light 816 may be processed and stored as the first imaging data. For example, the processing of the collected detection light 816 may comprise averaging the collected detection light 816. Subsequently, the selected one of the first imaging positions 330-1 (or a portion thereof) may be associated with the so generated first imaging data. The generated first imaging data may moreover be associated with the illumination positions 320 for which the detection light 816 is collected.

[0146] FIG. 10 shows an example of a scanning movement. The illumination light beam 816 is scanned (moved) along a saw tooth / shaped path across the first imaging positions 330-11, 330-12, 330-13, 330-21, 330-22, 330-23,. . . The imaging positions 330 are in aspect shown in FIG. 10 arranged in the shape of an array. The implementation-dependent prefactor c for setting the integration time dependent on the scanning frequency / scan (see above) is in the shown case equal to 1.

[0147] In an alternative example (not shown), the scanning movement may be sineshaped, wherein a quarter of the period of the sine-shaped scanning movement (phase < TT / 2) or half of the period of the sine-shaped scanning movement (phase < TT) may be used. In this alternative example, the implementation-dependent prefactor c for setting the integration time is, for the quarter period and the half period, equal to 2 and 1, respectively.

[0148] In one aspect of the disclosure, multiple ones of the first imaging data may be generated. The multiple ones of the first imaging data may, for example, result from different movements of the illumination light beam 812 associated with the selected one of the plurality of first imaging positions 330-1. The multiple ones of the first imaging data may be beneficial for training of ML models, since more information of the at least one sample 500 is available.

[0149] Generating the first imaging data by moving the illumination light beam 812 across at least the portion of illumination positions 320 results in more information about the at least one sample 500 being collected than in the case of collecting the detection light 816 for a single one of the illumination positions 320.

[0150] The analysing S120 (see FIG. 1) comprises analysing the first imaging data and identifying at least one region of interest (ROI). The analysis of the first imaging data to identify the at least one ROI may be done for example visually by a human, e.g., the pathologist, and / or by data processing, e.g., by means of the processing unit 713.-28-

[0151] The visual analysis of the first imaging data to identify the at least one RO I may comprise conversion of the first imaging data into a virtual H&E image. The virtual H&E image can then be assessed by the human. For example, the human may visually analyse the at least one ROI.

[0152] The conversion to the virtual H&E image may be based on the detected wavenumbers and corresponding ones of the intensities, detected for the plurality of first imaging positions 330-1. The conversion may further be based on determining one or more of an amount of cell nuclei, an amount of cell membranes, and / or an amount of cell bodies (cytoplasms) by comparing the detected intensities for the corresponding wavenumbers, such as for 2930 cmA-l (CH3-bond) and / or for 2845 cmA-l (CH2-bond).

[0153] The analysis of the first imaging data to identify the at least one ROI may be performed by means of a trained machine-learning (ML) model 714. The trained ML model 714 may receive as input data obtained the first imaging data. The input data to the trained ML model 714 may be the first imaging data generated for the plurality of first imaging positions 330-1. The trained ML model 714 identifies the at least one ROI based on the first imaging data.

[0154] The imaging S130 of the at least one sample 500 is performed by moving the illumination light beam 812, in a manner similar to the imaging SI 10, to a second plurality of second illumination positions 320-2.

[0155] Similarly to the first imaging positions 330-1, the plurality of second imaging positions 330-2 may be associated with the plurality of second illumination positions 320-2 (or second clusters of illumination positions 320-2), as shown in FIG. 3.

[0156] In one aspect, the nearest-neighbour distance of the second imaging positions 330-2 may be larger than the focal dimensions LI and L2 (focus size). The nearest-neighbour distances of the second imaging positions 320-2 may be, e.g., 2pm or 5 pm.

[0157] In another aspect, the plurality of second imaging positions 330-2 may each be associated with a single one of the plurality of second illumination positions 330-2. In this case, the nearest-neighbour distance of the second illumination positions 320-2 may be of a similar size or smaller size than the focal dimensions LI and L2. The plurality of second illumination positions 320-2 are illumination positions used to illuminate the at least one sample 500 for imaging the at least one ROI.-29-

[0158] The imaging 130 is analogous to the imaging SI 10 in that the imaging S130 shares the features of the imaging 110 described above, except that the imaging SI 30 relates to the plurality of second illumination positions 320-2, the plurality of second imaging positions 330-2, and the second imaging data.

[0159] The imaging S130 comprises illuminating the at least one sample 500 at the plurality of second illumination positions 320-2 with the illumination light beam 812. The imaging SI 30 further comprises collecting the detection light 816. The imaging S130 further comprises generating the second imaging data.

[0160] For any one of the plurality of second imaging positions 330-2, the second imaging data may comprise second imaging data elements associated with any one of the one or more wavenumbers (that are associated with the chemical markers). The second imaging data elements of the second imaging data may comprise intensities of the collected light for the corresponding wavenumber associated therewith. The generated second imaging data for any one of the plurality of second imaging positions 330-2 may be used to generate one or more second values. Based on the one or more second values and the corresponding one of the plurality of second imaging positions 330-2, a second image of the at least one sample 500 may be constructed. The image may be an in-depth image of the at least one ROI. The in-depth image may have a higher resolution than the overview image. The second image of the at least one ROI may have a resolution that is similar to or exceeds the resolution of H&E WSIs. The in-depth image of the at least one ROI enables the examination of the at least one sample 500 at a higher level of detail. The higher level of detail enables reassessing the identified at least one ROI. The reassessment based on the in-depth image enables improving the quality of diagnostic decisions.

[0161] Analysing S140 (see FIG. 1) comprises analysing the second imaging data. The analysis S140 of the second imaging data can be done, for example, visually or by data processing, e.g., by means of the processing unit 713.

[0162] The method may further comprise repeating S150 steps 120, 130, and 140. The method may thus comprise a further identification of further ROIs. The further ROIs may lie within the previously identified ROIs. The previously identified ROIs may be rescanned in a fashion analogous to S120 at a plurality of further illumination positions 320-3 (see FIG. 3). The further illumination positions 320-3 correspond to further imaging positions 330-3 (see FIG. 3). The further imaging positions 330-3 may have an nearest--30- neighbour distance which is smaller than the nearest-neighbour distance of the second imaging positions 330-2.

[0163] As mentioned above, SRS microscopy is capable of incorporating information, besides information relating to the C-H stretch region, about the at least one sample 500. The first imaging data and / or the second imaging data may be collected for the fingerprint region of the chemical markers. The chemical markers may be chosen according to a disease or condition of the at least one sample 500 that is to be assessed. The collected data about the chemical markers may be incorporated into ML models. For instance, if the chemical marker which is indicative for a disease is chosen additionally to the chemical markers from the C-H stretch region, as discussed above, the ML model may be enhanced (for classifying this disease) resulting in a more precise classification of the at least one sample 500. Assessing the at least one sample 500 based on both the fingerprint region and the C-H stretch region thus enables improved diagnostic capabilities.

[0164] In one aspect, the ML model is an artificial neural network (ANN). An example of the ANN is a Multilayer Perceptron (MLP) artificial neural network, also termed feedforward neural network (FNN). Another example of the ANN is a Convolutional Neural Network (CNN) or a Recurrent Neural Network (RNN). It will be understood that other types of ML models are also applicable.

[0165] The ANN comprises a plurality of layers. Each layer is associated with a rank / / , where n is, e.g., an integer number. The plurality of layers may comprise an input layer, one or more hidden layers, and an output layer. In one example, the input layer may be associated with the rank n = 1, the one or more hidden layers may be associated with the ranks n = 2, ... , N — 1, and the output layer may be associated with the rank n = N. The ANN further comprises a plurality of nodes. Any one of the plurality of nodes is associated with at least one input, at least one output, and an associated one of the plurality of layers.

[0166] The plurality of layers is ordered (or has an ordered structure). In the aforementioned example of associated ranks, a chosen node associated with one of the plurality layers of associated rank n = 2, ... , 1V receives the associated at least one input from ones of the plurality of nodes associated with a preceding one of the plurality of layers of associated rank n — 1. The at least one input and / or the at least one output of any one of the plurality of nodes result in a plurality of connections of the ANN between the nodes associated with the layer of rank n = 2, ... , 1V and the nodes associated with the layers of-31- rank n — 1. Any one of the plurality of connections may be associated with a connection weight.

[0167] Furthermore, in the example of associated ranks, any node of the input layer (having rank n = 1 may receive data as the associated at least one input. Examples of the data are training data (see below), the first imaging data, and / or the second imaging data. Moreover, any one node of the output layer having rank n = N may output a result which contributes to a calculated result generated by the ANN.

[0168] Any one of the plurality of nodes may have an activation function associated therewith. The activation function calculates the at least one output of the associated node based on the at least one input associated with the node and the connection weight associated with the corresponding one of the plurality of connections. For example, the connection weight may be factor, by which the corresponding at least one input is to be multiplied (weighted). The activation function takes the at least one input weighted by the connection weight as an argument (input to the activation function). The activation function may be a nonlinear activation function. The activation function includes ridge activation functions, radial activation functions, and fold activation functions. Examples of the ridge activation function are the ReLu (rectified linear unit) function, the Heaviside function, the linear function, and the Logistic function. Examples of the radial basis activation function are the Gaussian function, the multi quadratic function, the inverse multi quadratic function, and polyharmonic spline functions. Examples of the folding activation function are the mean, maximum, and minimum. Further examples of the activation function include the hyperbolic tangent function, the Soboleva modified hyperbolic step function, the GELU (Gaussian error linear unit) function, the binary step function, the Softplus function, the ELU (exponential line unit) function, the SELU (scaled exponential line unit), the Leaky ReLU (leaky rectified linear unit) function, the PreLu (parametric leaky rectified linear unit) function, the SiLU (Sigmoid linear unit, Sigmoid shrinkage, SiL, or Swish-1) function, Softmax function, Maxout function, but are not limited thereto.

[0169] The MLP is an example of a fully connected one of the ANN, in which, according to the afore-mentioned example of associated ranks, the chosen node associated with one of the plurality layers of associated rank n = 2, ... , N — 1 receives the associated input from each of the plurality of nodes associated with the preceding one of the plurality of layers of associated rank n — 1. In other words, each one of the plurality of nodes in the-32- layer of rank n = 2, ... , N — 1 is connected to each node in the subsequent layer of rank n + 1.

[0170] The purpose of the training process is to determine the value of the connection weight for each one of the plurality of connections. It will be understood that the ANN may be varied with respect to the number of layers as well as the number of nodes in each one of the layers. It will also be understood that different types of the ANN, e.g., the MLP, the CNNs, the FNN, but not limited thereto, differ in the setup of layers. Furthermore, the ANN may further comprise one or more convolutional layers, pooling layers, and / or normalization layers. These layers may also be varied with respect to the number nodes associated with the corresponding layer. Other parameters, for example the kind and size of the convolutional kernel, may differ between the one or more convolutional layers.

[0171] The training process may comprise a first iteration of the training and a second iteration of the training. The first iteration of the training process is carried out by providing first training data to the ANN. The first training data is provided to the input layer. The first training data may comprise, for example, the first imaging data and / or the second imaging data, as discussed with respect to the imaging SI 10 and / or the imaging S130. The first training data may also comprise image data, for example, but not limited to, H&E WSIs or pre-processed first imaging data, for example, virtual H&E images.

[0172] The first training data provided to the input layer is processed by the ANN and generates an output of the ANN. An error of the generated output of the ANN is determined by comparing the generated output to an expected result. The expected result is associated with the first training data and is provided in advance. For example, the first imaging data, the second imaging data, and / or the image data (e.g., the H&E WSIs) may be annotated by the pathologist to provide the expected result. The annotating includes, e.g., classifying regions in the first image, e.g. the overview image, constructed from the first imaging data, and / or the second image, e.g., the in-depth image, constructed from the second imaging data, as being in a diseased state or in a non-diseased state. The classifying may be performed manually or by interaction with a computer. The annotations associated with the first training data (or annotated first training data) thus provide the expected result. The annotations are usually referred to as “ground truth”.

[0173] For example, if the task of the ANN comprises classifying diseased tissue in an H&E image, the expected result is the classification of the first training data as diseased-33- tissue or non-diseased tissue according to the annotations. The output of the ANN is compared with the annotations (i.e., the expected result). The expected result provides a benchmark to evaluate the generated output of the ANN.

[0174] Based on the comparison of the generated output of the ANN and the expected result, the connection weights for each one of the plurality of connections are adjusted during the first iteration of the training. During the first iteration of the training, training parameters, such as, but not limited to, the amount of provided training data, or accuracy may be varied. This example of the first iteration of the training based on annotated training data is referred to as supervised learning. It will be understood that other methods, such as unsupervised learning, may also be applicable according to the present disclosure.

[0175] Once the first iteration of the training is completed, a final set of values for the connection weights between the layers is obtained.

[0176] In a further aspect of the invention, transfer learning is used to train the ANN. For transfer learning, a set of pre-trained weights of first one 714-1 of the trained ML model 714 (first trained ML model 714-1) is used for initialization of the connection weights of the ANN during training (or re-training) of the ANN. The transfer learning may be regarded as the second iteration of the training process for the training of the ANN, subsequent to the first iteration of the training described above. The second iteration of the training is applied to the ANN, in the manner described above. In the second iteration of the training, second annotated training data may be used for training (or re-training) the ANN. It will be understood that the concept of transfer learning is also applicable to the different types of artificial neural networks, for example the CNN, the FNN, and / or the RNN.

[0177] The concept of transfer learning enables using the knowledge obtained from the first iteration of the training, which results in the first trained ML model 714-1 to support the training of a second one 714-2 of the trained ML model 714 (second trained ML model 714-2). This use of the knowledge may result in the second training data being smaller in size compared to the first training data. The smaller size of the second training data reduces the required computational resources and / or the manual effort necessary to annotate the corresponding second training data.

[0178] In another aspect, the first training data and / or the second training data are pre-processed before the training process. For example, the pre-processing may comprise normalization of the first imaging data with respect to the laser power and / or signal-34- normalization for the chemical markers associated with the different vibrational states using a normalization target (or normalization sample) before the measurements of the first imaging data. The normalization target may comprise a collection of technical samples, e.g., plastic beads made of Polymethylmethacrylate (PMMA) or polystyrene or other samples used in the calibration of spontaneous Raman spectroscopy instruments. In a further example, the pre-processing comprises denoising and / or adjustment of image resolution of image data. The pre-processing of the first training data may comprise an adjustment of the image resolution. The pre-processing of the second training data may comprise the conversion of the obtained first imaging data to virtual H&E images.

[0179] FIG. 4 illustrates a flow chart of a method T. The method T illustrates in more detail the concept of transfer learning, described above, using the first training data and the second training data. The trained ANN (i.e., the second trained ML model 714-2) is then used to identify, e.g., the at least one ROI in the virtual H&E image or to assess the in-depth image gained from the second data, as discussed with respect to the method S (shown in FIG. 1) for imaging the at least one sample 500.

[0180] It will be understood that other optical imaging techniques than SRS microscopy may be used to obtain the virtual H&E images. Some of the other optical imaging techniques may require additional agents, such as fluorescent dyes, and / or additional ML-based data processing routines, to obtain the virtual H&E images. The other optical imaging techniques may comprise confocal reflectance microscopy, multiplexed stain-free microscopy (MUSE), photoacoustic imaging, multiphoton imaging, endoscopic in-vivo fluorescence imaging, radio frequency spectroscopy, fluorescence-based confocal laser scanning microscopy, light sheet microscopy, mid-infrared spectroscopy, and nonlinear microscopy.

[0181] The method of training an ANN for imaging the at least one sample 500, shown in FIG. 4, comprises providing T100 the first training data. The training data may comprise the annotated H&E WSIs 502A of training samples 500A as the first training data for the training of the first ML model. The annotation of the H&E WSIs 502A includes manually labelling by the pathologist, wherein the labelling indicates cancerous tissue or, more generally, diseased tissue. The annotating of the diseased tissue may comprise delimiting the diseased tissue and the non-diseased tissue.-35-

[0182] In one example, publicly available ones of the annotated H&E WSIs 502A were used as the first training data (see FIG. 6). The publicly available ones of the annotated H&E WSIs 502A were obtained from the CAMELYON16 and CAMELYON17 data sets (“https: / / camelyonl7.grand-challenge.org / Data / ”). The CAMELY0N16 and CAMELY0N17 data sets comprise training samples 500A of lymph node sections. The annotations of the CAMELY0N16 and CAMELY0N17 data sets provide the ground truth (i.e., the assessment of the images regarded as being correct). In this case, ground truth relates to metastases found in an individual one of the training samples 500A (e.g., the presence and / or the absence of metastases) and to a pN-stage of the individual one of the training samples 500A. The pN-stage provides a pathological lymph node classification of the individual one of the training samples 500A. The annotations of the training samples 500A were performed under supervision of expert pathologists.

[0183] The method may further comprise pre-processing T105 of the first training data. In the described case of using the CAMELY0N16 and CAMELY0N17 data sets, the first training data comprise the annotated H&E WSIs 502A provided in providing T 100 of the first training data. The pre-processing T105 may comprise adjusting of the image resolution of the annotated H&E WSIs 502A. For example, the adjustment may comprise a reduction of the image resolution of the annotated H&E WSIs 502A. The reduction of the image resolution may comprise, for example, line averaging, but is not limited thereto. It will be understood that other image processing techniques for adjusting image resolution may also be applicable.

[0184] In an example of the line averaging, a selected one of the H&E WSIs 502A is associated with an image grid 510, shown in FIG. 5B. The grid 510 shown in FIG. 5B is regular in shape; however, it will be appreciated that the grid 510 may be irregular in shape. The image grid 510 is composed of a plurality of image pixels 515 having an image pixel size L3, L4. The image pixel size (L3, L4) may correspond to a height and / or a width of the image pixel 515. As can be seen in FIG. 5B, a plurality of image pixel clusters 530 may be defined in the image grid 510. Any one of the plurality of image pixel clusters 515 comprises two or more image pixels 515 of the image grid 510. The two or more image pixels 515, forming one of the plurality of image pixels clusters 530, may be arranged adjacent to, or in proximity of each other. The image pixels clusters 530 shown in FIG. 5B have a square shape, i.e., having the same number of image pixels 515 in a vertical direction DI and a-36- horizontal direction D2; however, it will be appreciated that the image pixels clusters 530 may be of a non-square shape, e.g., a rectangular shape or a polygonal shape. The image pixel clusters 530 may have sixteen (four-by-four) image pixels 515 or eighty-one (nine-by- nine) image pixels 515. However, it will be appreciated that these are mere examples and that the image pixel clusters 530 may comprise other numbers of image pixels 515 than sixteen or eighty-one image pixels 515, e.g., less than sixteen / eighty-one or more than sixteen / eighty-one, e.g., five, twenty, or hundred image pixels 515. In a further aspect, some of the image pixels clusters 530 may comprise more or less image pixels 515 than other ones of the image pixels clusters 530.

[0185] In another aspect of the invention, the number of image pixels 515 of any one of the plurality of the image pixel clusters 530 corresponds to a portion of the training sample 500A, i.e., in the case of the training of the first ML model the lymph node sections, that has an extension of more than 5 pm. In yet another aspect, the number of image pixels 515 of any one of the plurality of the image pixel clusters 530 correspond to the portion of the training sample 500 A with the extension being larger than 10 pm. Each image pixel 515 is associated with a corresponding image pixel of the underlying annotated H&E WSI 502A. An averaging of two or more image pixels 515 is then performed for at least one of the plurality of image pixel clusters 530 to generate a pre-processing value associated with the at least one of the plurality of image pixel clusters. The pre-processing T105 of the H&E WSIs 502A may result in pre-processed ones of the first training data having a lower- resolution.

[0186] In FIGS. 5A to 5C, the pre-processing T105 is illustrated for reducing the image resolution of the H&E WSIs 502A. FIG. 5 A shows the H&E WSI 502A. FIG. 5 A further shows in the inset a magnified portion 505 of the H&E WSI 502A. FIG. 5B shows a schematic illustration of the image grid 510. The averaging of the values of the image pixels 515 may be performed, for example, along or parallel to an averaging path 535. The averaging path 535 may, for example, be directed horizontally, vertically, or diagonally. The averaging path 535 may, in another example, follow a non-straight path. FIG. 5C shows a comparison between the magnified portion 505 and the resolution-reduced magnified portion 505’.

[0187] The line averaging technique, as described above, integrates information along the averaging path 535. Pre-processing the first training data (for example, annotated-37-H&E WSIs 502A of the training samples 500A) of the first ML model, i.e., reducing the image resolution, enables adapting the resolution of the first training data with respect to the resolution of the second training data of the second ML model (i.e. resolution of the virtual H&E images). The adapting improves the performance of the resulting trained ML model 714. The improvement enables a more precise classification of the at least one sample 500.

[0188] In one aspect of the disclosure, the averaging of the plurality of image pixels 515 is chosen to resemble a possible movement (scan path) 331, 332, 333 of the illumination light beam 812 in the imaging T120. Ones of the image pixels 515 illuminated by illumination beam 812 along the scan path are averaged. To provide more training data (data augmentation), several scan paths may be taken into account to create additional pre- processed data sets. The additional pre-processed data sets may have a decreased resolution. For further data augmentation, the origin of the grid 510 may be shifted in a vertical direction DI and a horizontal direction D2 before averaging, to form more additional pre-processed training datasets. The shifting may be repeated with the shifting of the origin of the grid 510 by a different number of image pixels 515 D2’ .

[0189] The method further comprises training T110 the first ML model. The training T110 comprises training the first ML model on the first training data or on the pre-processed first training data. In the case described above, the first ML model is trained on the annotated H&E WSIs 502A or on the pre-processed annotated H&E WSIs 502A. In one example, the ML model is an artificial neural network (ANN). In one aspect of the disclosure, the ANN may be a CNN based on a modified ResNet50 using a rectified linear unit (ReLu) as the activation function, the optimizer Adam as optimizer, and the Binary Cross Entropy loss as loss function. During validation of this ResNet50, a learning rate, a batch size, a threshold for classification as cancerous tissue, and methods as well as numbers of augmentations applied to the training data were tuned. It will be understood that other ML models may also be applicable, such as models from the domains of supervised learning, unsupervised learning, or reinforcement learning.

[0190] The method further comprises a imaging T120 a sample 500 of interest (or samples 500 of interest) to generate second training data. In the above-described example of training the ANN, the generating T120 the second training data may comprise generating the virtual H&E images. In one aspect, the second training data may be obtained by using SRS microscopy and performing the imaging SI 10, as discussed with respect to FIG. 1. The-38- image resolution of the generated virtual H&E images may be similar to the image resolution of the resolution-reduced annotated H&E WSIs used in the training T110.

[0191] The method further comprises annotating T 130 the generated second training data. In the above-described example of training the ANN, the annotating T130 of the generated second training data may comprise annotating the generated virtual H&E images.

[0192] The method further comprises applying T140 the transfer learning to the first trained ML model using the annotated second training data. The pre-trained connection weights of the first trained ML model of the training T110 are used to initialize the connection weights of the second ML model. After the training of the second ML model has been completed, the connection weights of the second ML model may be adapted based on the second training data, e.g., the annotated virtual H&E images. The trained (second) ML model 714 can be used, for example, in the method S for analysing the first imaging data, e.g., for identifying ROIs of the at least one sample 500, as discussed with respect to FIG. 1. The trained (second) ML model 714 can further be used, for example, in the method S for analysing the second imaging data for assessing the at least one sample 500, as discussed above.

[0193] FIG. 6 shows a pre-processed one of the H&E WSI 502A from the CAMELYON16 and CAMELYON17 datasets that has been reduced in its resolution. On the left side of FIG. 6, the ground truth that has been derived from the annotation is overlayed on the H&E WSI 502A. The pre-processed H&E WSI 502A is split into different tiles containing a multiple of reduced resolution pixels. The overlay shows the percentage of pixels in a tile that contain cancerous tissue. The right side of FIG. 6 shows the same H&E image 502A analysed by means of the first trained ML model and overlaid with predictions of cancerous tissues in the different regions.

[0194] FIG. 7 illustrates a scenario in which the method S, as discussed with respect to FIG. 1, is applied. The scenario relates to a use case in which a surgeon 702 classifies tissue (i.e., the sample 500), which potentially results in removing the tissue from a patient 703 during surgery. The tissue to be classified may have been acquired during a biopsy or the surgery. The surgery may be, for example, but is not limited to, a lumpectomy (e.g., a breast conserving surgery). In case the tissue has been classified as containing the diseased tissue, the surgeon 702 removes the diseased tissue while leaving as much healthy tissue as possible. The method according to the present disclosure enables the surgeon 702 to-39- differentiate between healthy or non-diseased tissue and diseased, e.g., cancerous, tissue, contained in the classified tissue. The removed tissue is imaged with an optical imaging system 710 to classify the removed tissue. The optical imaging system 710 comprises the SRS microscope (see above). The SRS microscope provides the virtual H&E image, e.g., the H&E overview image, of the removed tissue. The overview image may then be displayed on a display device 732 in the operation room (OR). The display device 732 may comprise a (flat) screen, a computer monitor / screen, a tablet device, or a smartphone. The overview image may also be displayed using a graphical user interface (GUI) 850 on the display device 732. The GUI 850 may provide, for example, interactive control elements for magnifying portions of the overview image. The overview image can be assessed by the surgeon 702 and / or the pathologist 704 to assess whether an excision of tissue from the patient is necessary. In one aspect, the excision may take place in proximity to a location from which the classified tissue has been removed. In another aspect, the excision may take place around a location of the body of the patient, e.g., on the skin, that has been imaged. The excision removes tissue from the patient based on the classification of the classified tissue, e.g., when the classified tissue indicates that it is likely that further diseased tissue is present in the patient (e.g., the tumour has not been entirely removed during the first excision), or when a portion of the skin of the patient has been classified as being diseased. Since the imaging process of the tissue, e.g., the removed tissue, is fast and the assessment of the tissue can be performed during surgery in the OR (intraoperative assessment), the patient 703 remains under anaesthesia until the diagnosis of the pathologist 704 is available. Hence, there is no need for scheduling subsequent surgeries, as may be the case for imaging processes using manual (traditional) H&E staining procedures. This reduces stress on the patient 703, risks of surgical infections from the re-excision, and saves costs due to more efficient use of medical infrastructure.

[0195] In another aspect of the invention, the pathologist 704 identifies and marks one or more ROIs in the overview image during the surgery. The one or more RO Is are subsequently reimaged by the optical imaging system 710 at a higher resolution to generate the in-depth images, as described above with respect to the method S. The pathologist 704 can then reassess the marked regions. This increases the assessment quality based on the classification of the at least one sample 500 and leads to better diagnoses.-40-

[0196] In another aspect of the invention, the obtained virtual H&E images of the removed tissue are analysed not only visually by the pathologist 704 but also via the trained ML model 714. The trained ML model 714 outputs one or more identified ROIs. The one or more ROIs are subsequently reimaged by the optical imaging system 710 at a higher resolution to generate the in-depth images, as described above with respect to the method S. These in-depth images are then already available when the pathologist 704 visually analyses the sample. The trained ML model 714 can thus assist the pathologist 704 in the decisionmaking and may also save the pathologist waiting time for acquiring in-depth images in the ROIs. This increases the assessment quality regarding the tissue to be analysed and leads to better diagnoses, since pathological analyses are subjective and prone to errors, for example, due to exhaustion and overload of the pathologist 704.

[0197] In another aspect of the invention, the obtained virtual H&E images of the tissue to be analysed are provided to the pathologist 704 remotely during surgery, for example, via a digital connection. The pathologist 704 can communicate his / her assessment / diagnosis to the surgeon 702, for example, by telephone or other communication means.

[0198] FIG. 9 shows one example of the overview image gained from the first imaging data represented as a virtual H&E image with the first nearest-neighbour distance having a value of 20 pm. This example, illustrated in FIG. 9, was acquired by scanning a surface area of the at least one sample 500 of over 10 mm x 10 mm in about 40 s. For the overview image shown in FIG. 9, the first imaging data were converted to a grayscale.REFERENCE NUMERALSS methodT method310 detector grid311 origin of detector grid315 detector pixel320 illumination positions330-1, 330-2 imaging positions331 horizontal line332 vertical line333 diagonal lineLI, L2 detector pixel sizeDI vertical directionD2 horizontal direction500 sample500A training sample502A raw, pre-processed, or annotated H&E WSI of a training sample505, 505’ portion of an H&E WSI image510 image grid515 image pixel530 image pixel clustersL3, L4 image pixel size702 surgeon703 patient704 pathologist710 optical imaging system712 illumination module713 processing unit714 trained machine-learning (ML) model732 display device810 light source812 illumination light beam813 scanner814 illumination objective lens815 detection obj ective lens816 detection light820 control unit830 detection module832 spectrometer unit834 pixel detector840 memory850 graphical user interface (GUI)

Claims

-42-Claims1. A method (S) for imaging at least one sample (500), comprising:- providing (S100) an illumination light beam (812) comprising light of one or more illumination wavelengths for illuminating the at least one sample (500);- imaging (SI 10) the at least one sample (500) at a plurality of first imaging positions (330-1), the imaging (SI 10) comprising(i) illuminating, at a plurality of first illumination positions (320-1), the at least one sample (500) with the illumination light beam (812);(ii) collecting first detection light (816) from the plurality of first illumination positions (320-1);(iii) generating, from the collected detection light (816), first imaging data associated with the plurality of first imaging positions (330-1), wherein, for at least one of the plurality of first imaging positions (330-1), the illumination light beam (812) is moved across several ones of the plurality of first illumination positions (320-1) during the collecting of first detection light (816) to generate the first imaging data, associated with the at least one of the plurality of imaging positions (330-1), from the several ones of the plurality of first illumination positions (320-1);- analysing (S120) the first imaging data with respect to at least one detection wavelength, to identify at least one region of interest (RO I);- imaging (SI 30) the at least one ROI at a plurality of second imaging positions (320-2), the imaging (SI 30) comprising(iv) illuminating, at a plurality of second illumination position (320-2), the at least one sample (500) with the illumination light beam (812);(v) collecting second detection light (816) from the plurality of second illumination position (320-2);(vi) generating second imaging data associated with the plurality of second positions; and- analysing (S140) the second imaging data with respect to the at least one emission wavelength.-43-2. The method (S) of claim 1, wherein the collecting of the detection light comprises moving the illumination beam (812) among at least some of the plurality of illumination positions (320-1, 320-2, 330-3).

3. The method (S) according to claim 1, wherein the analysing (SI 20) of the first imaging data comprises inputting the first imaging data into a trained machinelearning model (714).

4. The method of claim 3, wherein the method further comprises training the machinelearning model (714) and pre-processing training data used in the training.

5. The method of claim 4 wherein the training data are image data and the preprocessing comprises reducing a resolution of the image data.

6. The method (S) according to any one of claims 1 to 5, wherein the analysing of the first imaging data comprises detecting at least one chemical marker.

7. The method (S) according to claim 6, wherein the detecting of the at least one chemical marker comprises Stimulated Raman Scattering microscopy or Stimulated Raman Photothermal microscopy.

8. The method (S) according to any one of the preceding claims, wherein the analysing of the second imaging data comprises inputting the second imaging data into a / the trained machine-learning model (714).

9. The method (S) according to any one of the preceding claims, wherein a first nearest- neighbour distance of the plurality of first positions (320-1) is larger than a second nearest-neighbour distance of the plurality of second positions (320-2).

10. The method (S) according to claim 9, wherein the first nearest-neighbour distance is larger than 5 pm and the second nearest-neighbour distance is equal to or smaller than 5 pm.-44-11. The method (S) according to any one of the preceding claims, further comprising user-based identifying, based on the first imaging data, of at least one further ROI.

12. The method (S) according to any one of the preceding claims, further comprising performing the imaging (SI 30) on the at least one further ROI for a plurality of further second positions.

13. The method (S) according to any one of the preceding claims, further comprising storing the first imaging data generated for the plurality of first positions as first image data, and / or the second imaging data generated for the plurality of second positions as second image data.

14. The method of claim 13, further comprising displaying the first image data and / or the second image data.

15. The method (S) of claim 14, further comprising providing a graphical user-interface for visual analysis the first image data and / or the second image data.

16. The method (S) of any one of claims 1 to 14, wherein the generating of the first imaging data comprises integrating at least one electrical signal generated from the first detection light (816).

17. The method (S) of claim 16, wherein an integration time of the integrating of the at least one electrical signal is adapted to a scanning frequency / scan of the moving of the illumination light beam (812) across the several ones of the plurality of first illumination positions (320-1).

18. The method (S) of claim 17, wherein the integration time of the integrating of the at least one electrical signal is set such that a predefined signal-to-noise ratio is achieved.-45-19. The method of any one of claims 1 to 18, wherein the method further comprises image mosaicking.

20. An optical imaging system (710) for imaging at least one sample (500), comprising an illumination module (712) configured to provide an illumination light beam (812) for illuminating the at least one sample (500) at a plurality of illumination positions (320), the illumination light beam (812) comprising light of one or more illumination wavelengths; a control unit (820) configured to move the illumination light beam (812) to the plurality of illumination positions (320-1, 320-2, 320-3); a detection module (830) configured to collect, for any one of the plurality of illumination positions (320), detection light (816) from the plurality of illumination positions (320-1, 320-2, 320-3) and to convert the detection light (816) into data; a memory (840) configured to store the data received from the detection module (830) and associated with the corresponding one of the plurality of imaging positions (330); and at least one processing unit (713) configured to analyse the data with respect to at least one detection wavelength, to identify at least one region of interest (ROI), wherein the control unit (820) is configured, for at least one of the plurality of imaging positions (330-1), to move, during collecting of the detection light (816) by the detection module (830), the illumination light beam (812) across several ones of the plurality of illumination positions (320) to generate the data, associated with the at least one of the plurality of imaging positions (330-1), from the several ones of the plurality of illumination positions (320-1).

21. The optical imaging system (710) according to claim 20, wherein the detection module (830) comprises at least one spectrometer unit (832) for detecting a plurality of signature wavelengths indicative of a plurality of chemical markers.

22. The optical imaging system (710) according to claim 20 or 21, configured to input the data into at least one trained machine-learning model (714).-46-23. The optical imaging system (710) according to any one of claims 20 to 22, further comprising at least one display device (732), in communication with the at least one processing unit (713) and configured to display an image based on the data, the image optionally indicating the at least one ROI.

24. The optical imaging system (710) according to claim 23, wherein the optical imaging system (710) is further configured to provide a graphical user interface (850) for visually analysing the data by means of the at least one display device (732).

25. The optical imaging system (710) according to any one of claims 20 to 24, further comprising a lock-in amplifier (713).

26. The optical imaging system (710) according to any one of claims 20 to 25, further comprising a stage (855) for moving the sample (500).

27. A computer program product comprising instructions which, when the program is executed by a processing unit (713) of the optical imaging system (710) of claim 13, cause the optical imaging system (710) to perform the method of any one of claims 1 to 19.

28. A computer-readable storage medium having stored thereon the computer program product of claim 27.

Images