AgInS quantum dot bio-imaging image enhancement identification system and method

Abstract

The application relates to the technical field of biomedical imaging data processing, and discloses an AgInS quantum dot biological imaging image enhancement and recognition system and method. The system comprises an excitation light source, a gated camera, a time sequence controller and a data processing unit. The method comprises the following steps: receiving image frames under different delay time windows to construct a three-dimensional space-time data matrix; orthogonalizing a pre-calibrated long-short-life prior vector to generate an initial target base, and performing inner product calculation on the initial target base and a pixel time sequence vector to generate an initial target weight matrix; performing segmentation and corrosion treatment on the initial target weight matrix to extract a high-accuracy background mask, performing reverse mapping to obtain a dynamic background base, and performing orthogonalization and inner product calculation on the dynamic background base to obtain a final target weight matrix; and performing numerical mapping on the final target weight matrix and performing connected region analysis to output a result. The application realizes real-time adaptive updating of background reference features, and improves target contrast and target recognition accuracy in a complex living body environment.

Description

Technical Field

[0001] This invention relates to the field of biomedical imaging data processing technology, specifically to an AgInS quantum dot bioimaging image enhancement and recognition system and method. Background Technology

[0002] Fluorescence imaging technology has wide applications in biomedical diagnosis and research. AgInS quantum dots, due to their long fluorescence lifetime, are often used as contrast probes for in vivo targeted imaging. During in vivo optical imaging, the excitation source not only excites the quantum dots aggregated in the target region but also excites endogenous tissues, generating endogenous autofluorescence. Compared to quantum dots, the endogenous autofluorescence lifetime of living tissues is typically shorter. Existing techniques usually utilize the temporal differences in fluorescence decay lifetimes of luminescent materials, by setting a fixed time delay acquisition window or establishing a fixed mathematical projection model based on pre-calibrated static prior data, to filter out short-lived background fluorescence and extract long-lived target signals.

[0003] However, the living organism's interior is a complex and dynamically changing physical and physiological environment. Changes in local microcirculation, metabolic states, and organ physiological movements all cause dynamic fluctuations in the decay characteristics of endogenous autofluorescence over time and space. When using conventional fixed delay windows or static models for signal separation, the preset background reference cannot be updated synchronously with changes in the actual living environment. When the actual background signal deviates from the static prior model, the fixed algorithm cannot accurately align and cancel the dynamically changing interference components during the calculation process, resulting in significant residual background noise in the separated and reconstructed data. In deep tissues or imaging areas with strong tissue autofluorescence, this residual background can partially submerge the target signal, resulting in low contrast and blurred target boundaries in the final output image, failing to meet the practical application requirements for high-accuracy image segmentation and lesion identification. Summary of the Invention

[0004] The technical problem solved by this invention is that existing quantum dot in vivo bioimaging is easily interfered with by background signals such as endogenous tissue autofluorescence, and the in vivo background fluorescence characteristics change dynamically with time and physiological state. It is difficult to achieve accurate separation of target signals by using a fixed prior model, resulting in low accuracy of target signal recognition and low image contrast.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: The first aspect of this invention provides a method for enhancing and recognizing bio-imaging images using AgInS quantum dots, comprising the following steps: Two-dimensional image frames acquired under different delay time windows are received, and the two-dimensional image frames are stacked sequentially along the time dimension to construct a three-dimensional spatiotemporal data matrix; Load the pre-calibrated AgInS quantum dot long-lived standard time decay vector and the in vivo endogenous autofluorescence short-lived initial prior vector; perform orthogonalization operation on the long-lived standard time decay vector and the short-lived initial prior vector to generate the initial target substrate; Extract the temporal vector corresponding to each spatial pixel in the three-dimensional spatiotemporal data matrix, and calculate the inner product between the temporal vector and the initial targeting basis to generate the initial targeting weight matrix; The initial target weight matrix is ​​subjected to global adaptive segmentation and morphological processing to extract a binarized high-confidence background mask; The high-confidence background mask is reverse-mapped to the three-dimensional spatiotemporal data matrix, the temporal vector corresponding to the mask coordinates is extracted, and the arithmetic mean is performed to generate a dynamic background basis vector characterizing the current physiological interference features. The orthogonalization operation is re-performed by replacing the short-lifetime initial prior vector with the dynamic background basis vector to generate a secondary targeting basis; Perform an inner product projection operation between the temporal vector corresponding to the full field of view pixels and the secondary targeting basis to obtain the final targeting weight matrix; The final target weight matrix is ​​numerically mapped and converted into a two-dimensional enhanced grayscale image. Connectivity analysis is then performed on the two-dimensional enhanced grayscale image to output the recognition result.

[0006] The innovative principle of the above technical solution lies in utilizing the difference in temporal decay characteristics between long-lifetime AgInS quantum dot fluorescence and short-lifetime intrinsic autofluorescence to extract pixel-level temporal evolution features by constructing a three-dimensional spatiotemporal data matrix. First, an initial orthogonal space is constructed using pre-calibrated prior data, and signal components highly correlated with the background are stripped to achieve preliminary separation. Then, high-confidence background regions are extracted in the spatial domain, and a dynamic background substrate reflecting the real-time interference state is extracted and reconstructed from the currently acquired data. Finally, a secondary orthogonal projection operation is performed based on the updated dynamic background substrate. This process achieves real-time adaptive updating of background interference features, overcomes model bias caused by static background calibration, and effectively separates residual background interference in complex living environments.

[0007] Further, the orthogonalization operation performed on the long-lifetime standard time decay vector and the short-lifetime initial prior vector to generate an initial targeting basis includes: The short-lived initial prior vector is subjected to L2 norm calculation to generate a normalized background basis vector. The Gram-Schmidt orthogonalization method is used to calculate the projection component of the long-lived standard time decay vector along the direction of the normalized background basis vector to extract orthogonal residual vectors. The orthogonal residual vectors are then subjected to L2 norm normalization to generate the initial target basis. By calculating and stripping the projection component using the Gram-Schmidt orthogonalization method, background direction components can be filtered out from the target temporal features, ensuring that the generated basis direction is orthogonal to the interfering background and providing a specific response direction for subsequent inner product operations.

[0008] Further, the step of calculating the inner product between the temporal vector and the initial target basis to generate the initial target weight matrix includes: Calculate the inner product between the temporal vector and the initial target basis to obtain the initial target weight scalar at the spatial coordinates; perform non-negative truncation on the initial target weight scalar. If the initial target weight scalar is less than zero, assign it a value of zero; if the initial target weight scalar is greater than or equal to zero, retain its original value to obtain the initial target weight value; arrange the initial target weight values ​​corresponding to each spatial position in two dimensions according to the original spatial coordinates to generate the initial target weight matrix.

[0009] Further, the step of performing global adaptive segmentation and morphological processing on the initial targeting weight matrix to extract a binarized high-confidence background mask includes: The Otsu method is used to determine a global adaptive segmentation threshold based on the inter-class variance of foreground and background pixels in the initial target weight matrix. A binarized target region mask is generated from the initial target weight matrix according to the global adaptive segmentation threshold. Boolean inversion is performed on the target region mask to obtain a preliminary background region. A preset morphological structuring element is used to perform an erosion operation on the preliminary background region to generate the high-confidence background mask. The morphological erosion operation shrinks the preliminary background region inward, eliminating transitional mixed pixels near the target boundary, thereby extracting high-confidence background samples far from the target center and avoiding interference from long-lived target signals in subsequent background modeling steps.

[0010] Furthermore, after generating the target region mask, the method further includes: The percentage of pixels with a value of 1 in the target region mask is counted. If the percentage of pixels is higher than a preset upper threshold, the step of generating the dynamic background basis vector is not performed, and the short-lived initial prior vector continues to be used to generate the secondary target basis. If the percentage of pixels is lower than a preset lower threshold, an empty recognition result is directly output.

[0011] Further, the step of inversely mapping the high-confidence background mask to the three-dimensional spatiotemporal data matrix, extracting the temporal vector corresponding to the mask coordinates, and performing an arithmetic mean operation to generate a dynamic background basis vector characterizing the current physiological interference features includes: The number of effective background pixels in the high-confidence background mask is counted. When the number of effective background pixels is not less than a preset safety threshold, spatial coordinates satisfying a mask value of 1 are addressed in the two-dimensional plane, and temporal vectors corresponding to each spatial coordinate are extracted from the three-dimensional spatiotemporal data matrix. The extracted temporal vectors are accumulated element-wise along the time dimension, and the accumulated result corresponding to each time node is divided by the number of effective background pixels to obtain the dynamic background basis vector. Using the temporal feature mean of multiple background pixels in the entire domain as a new background reference can reflect the current actual environment while canceling single-point measurement noise, ensuring the numerical stability of the basis update process.

[0012] Further, the step of replacing the short-lived initial prior vector with the dynamic background basis vector and re-performing the orthogonalization operation to generate a secondary targeting basis includes: The dynamic background basis vector is normalized using the L2 norm to obtain a normalized dynamic background vector; the projection component of the long-lifetime standard time decay vector in the direction of the normalized dynamic background vector is calculated using the Gram-Schmidt orthogonalization method to obtain the orthogonal residual vector; the orthogonal residual vector is normalized using the L2 norm to obtain the secondary targeting basis.

[0013] Further, the step of performing an inner product projection operation between the temporal vector corresponding to the full field of view pixels and the secondary targeting basis to obtain the final targeting weight matrix includes: Calculate the inner product between the temporal vector corresponding to each spatial coordinate and the secondary targeting basis to obtain the secondary projection weight scalar; perform boundary constraint processing on the secondary projection weight scalar, setting it to zero when the secondary projection weight scalar is less than zero; retain its original value when the secondary projection weight scalar is greater than or equal to zero to obtain the final targeting weight value; arrange all the final targeting weight values ​​in a two-dimensional array according to the corresponding positions of each spatial coordinate to generate the final targeting weight matrix.

[0014] Further, the final target weight matrix is ​​numerically mapped to convert it into a two-dimensional enhanced grayscale image, and connected component analysis is performed on the two-dimensional enhanced grayscale image to output the recognition result, including: The maximum and minimum weight values ​​in the final target weight matrix are extracted. A linear stretching mapping relationship is established based on the maximum and minimum weight values ​​to map floating-point weight values ​​to the grayscale range corresponding to a preset bit depth, generating the two-dimensional enhanced grayscale image. The two-dimensional enhanced grayscale image is segmented based on a preset activation threshold, and activated pixels that meet the threshold conditions are selected. Connected regions of the activated pixels are merged according to the adjacency relationship between pixels. The area of ​​the merged connected regions is statistically analyzed, and connected regions with an area less than a preset lower area threshold are removed. Connected regions with an area not less than the preset lower area threshold are retained as valid connected regions. The geometric centroid coordinates are calculated based on the spatial coordinates of all valid pixels within the valid connected regions, and the outer contour boundary is extracted to generate contour parameters. The recognition result is then output.

[0015] A second aspect of the present invention provides an AgInS quantum dot bioimaging image enhancement and recognition system, comprising: Excitation light source, used to emit periodic excitation light pulses towards the target area of ​​a living organism; A gated camera is positioned on the return optical path of the excitation light source to receive the fluorescence signal generated by the target tissue after excitation, and to acquire images of the fluorescence signal to output a two-dimensional image frame. A timing controller is communicatively connected to the excitation light source and the gated camera, respectively, and is used to output a synchronous trigger signal to control the light pulse emission period of the excitation light source, and to control the gated camera to perform shutter opening and closing operations within different delay time windows after the excitation light pulse; The data processing unit is communicatively connected to the gated camera and the timing controller, respectively, and is used to execute the above-described AgInS quantum dot bioimaging image enhancement and recognition method.

[0016] Based on the above-described system platform, the timing controller coordinates the excitation of the light source pulse and the delay timing of the gated camera shutter to complete the time-step sampling operation; the data processing unit realizes matrix orthogonal stripping and dynamic background reconstruction operations at the hardware communication level, thereby completing the spatial positioning of the target signal at the level of physical acquisition and logical algorithm combination.

[0017] This invention provides an AgInS quantum dot bioimaging image enhancement and recognition system and method. It has the following beneficial effects: 1. This invention extracts a high-confidence background mask by segmenting and morphologically processing the initial target weight matrix, then maps this mask back to a three-dimensional spatiotemporal data matrix. The corresponding temporal vectors are extracted, and their average values ​​are calculated to generate a dynamic background basis. Finally, orthogonalization is re-performed based on this dynamic background basis, and a secondary projection is completed. This technical solution achieves real-time adaptive updating of the background reference signal, enabling synchronous adjustment of the background separation direction according to physiological changes in the endogenous environment of the living organism. It overcomes the computational bias caused by using a single fixed prior calibration model, thereby accurately separating and filtering out residual background interference in the real-time environment.

[0018] 2. This invention utilizes the temporal difference in fluorescence decay lifetime between AgInS quantum dots and living tissue. It loads a long-lifetime standard time decay vector and a short-lifetime initial prior vector, and extracts an orthogonal residual vector through orthogonalization to generate a targeting substrate. This residual vector is then multiplied by a temporal feature vector constructed using different delay windows. This process mathematically decouples the mixed luminescence signals directionally, directly removing components related to the autofluorescence background during temporal evolution, effectively suppressing the target signal overwhelmance phenomenon caused by strong in vivo autofluorescence.

[0019] 3. When extracting background reference samples, this invention inverts the segmented target region mask and performs erosion operations on the obtained region using morphological structuring elements to generate a high-confidence background mask. The morphological erosion operation causes the selected background region to shrink inwards spatially, actively eliminating transitional mixed pixels near the boundary of the luminescent target. This avoids the erroneous inclusion of long-lived target components into the background reference vector, thereby preventing the target signal from being excessively canceled or incorrectly subtracted during subsequent secondary signal separation.

[0020] 4. This invention introduces a non-negative truncation process when calculating the target weight scalar, filtering out meaningless negative values ​​caused by hardware acquisition fluctuations or physical noise. Simultaneously, before entering the dynamic background extraction stage, it performs conditional checks on the pixel ratio of the target mask and the number of effective pixels in the high-confidence background. When near-full target coverage is detected in the field of view or the number of effective background samples falls below the safety lower limit, the dynamic update is automatically terminated and the initial prior vector is used again. These mechanisms avoid division-by-zero anomalies and computational divergence caused by missing background samples, ensuring stable operation of the system under various extreme imaging conditions.

[0021] 5. This invention performs a linear stretching mapping of the final target weight matrix to its maximum and minimum values, converting it into a two-dimensional enhanced grayscale image adapted to a conventional bit depth. Subsequently, pixels are merged based on threshold activation and connected component analysis, isolated noise points with areas below the lower limit are removed, and the geometric centroid coordinates and outer contour parameters of the retained regions are calculated. This scheme accurately restores the orthogonal projection results of high-dimensional temporal data to a two-dimensional physical space coordinate system, providing downstream medical imaging workstations with intuitive observation images and quantified lesion localization data. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the image enhancement and recognition system of the present invention; Figure 2 This is a schematic flowchart of the image enhancement and recognition method of the present invention; Figure 3 This is a schematic diagram illustrating the principle of the three-dimensional spatiotemporal data matrix construction process of the present invention. Figure 4 This is a schematic diagram of the operational logic of the initial target substrate generation process of the present invention; Figure 5 This is a schematic diagram illustrating the initial target weight matrix calculation process of the present invention. Figure 6 This is a schematic diagram illustrating the high-reliability background mask extraction process of the present invention. Figure 7 This is a schematic diagram of the dynamic background substrate resampling process of the present invention; Figure 8 This is a schematic diagram illustrating the principle of the quadratic orthogonalization and final target weight matrix calculation process of the present invention. Figure 9 This is a schematic diagram of the target recognition result output process of the present invention; Figure 10 This is a comparison chart of the target identification results of multiple live tumor samples at different processing stages according to the present invention; Figure 11 This is a quantitative comparison chart of the credit-to-background ratio under different algorithms of the present invention; Figure 12 This is a target classification subject operating characteristic curve of the present invention.

[0023] Among them, 10 is the excitation light source; 20 is the gated camera; 30 is the timing controller; and 40 is the data processing unit. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1 The present invention provides an AgInS quantum dot bioimaging image enhancement and recognition system, which may include: an excitation light source 10, a gated camera 20, a timing controller 30, and a data processing unit 40.

[0026] The excitation light source 10 is used to emit periodic excitation light pulses to the target area of ​​a living organism.

[0027] The gated camera 20 is positioned on the return optical path of the excitation source 10 to receive the fluorescence signal generated by the target tissue after excitation and to acquire images of the fluorescence signal.

[0028] The timing controller 30 is communicatively connected to the excitation light source 10 and the gated camera 20, respectively, and is used to output synchronous trigger signals to coordinate the timing of each hardware unit. The timing controller 30 is used to control the light pulse emission period of the excitation light source 10 and to control the gated camera 20 to perform shutter opening and closing operations within different delay time windows after the excitation light pulse.

[0029] The data processing unit 40 is communicatively connected to the gated camera 20 and the timing controller 30, respectively. When an abnormal number of received frames, out-of-bounds acquisition parameters, or the need for re-acquisition is detected, the data processing unit 40 sends a re-acquisition command or parameter adjustment command to the timing controller 30. The data processing unit 40 can also be connected to a display interface or a downstream medical processing workstation via an external communication interface.

[0030] The data processing unit 40 is used to receive image data output by the gated camera 20, perform calculations on the image data, and output enhanced grayscale images, connected region coordinates, centroid coordinates, and contour parameters.

[0031] See attached document Figure 2 This invention provides a method for enhancing and recognizing bioimaging images using AgInS quantum dots, comprising the following steps: S10, the data processing unit 40 receives two-dimensional image frames acquired by the gated camera 20 under different delay time windows, and stacks the two-dimensional image frames sequentially along the time dimension to construct a three-dimensional spatiotemporal data matrix.

[0032] S20, the data processing unit 40 loads the preset data obtained from offline calibration. The preset data includes the AgInS quantum dot long-lived standard time decay vector and the in vivo endogenous autofluorescence short-lived initial prior vector. The data processing unit 40 performs orthogonalization operation on the long-lived standard time decay vector and the short-lived initial prior vector to generate the initial target substrate.

[0033] S30, the data processing unit 40 traverses each spatial pixel in the three-dimensional spatiotemporal data matrix, extracts the temporal vector corresponding to each spatial pixel, and calculates the inner product between the temporal vector and the initial target basis to obtain the initial AgInS signal weight corresponding to each pixel, and then generates the initial target weight matrix.

[0034] S40, the data processing unit 40 performs global adaptive segmentation and morphological processing on the initial target weight matrix to strip off transition pixels and extract a binarized high-confidence background mask.

[0035] S50, the data processing unit 40 reverse maps the high-confidence background mask to a three-dimensional spatiotemporal data matrix, extracts the temporal vector corresponding to the mask coordinates, and performs an arithmetic mean operation on the extracted temporal vectors to generate a dynamic background substrate that characterizes the current physiological interference features.

[0036] S60, the data processing unit 40 replaces the initial prior vector of short-lived endogenous autofluorescence in living organism used for orthogonalization operation in step S20 with the dynamic background basis vector, and re-executes the orthogonalization operation to generate a secondary targeting basis; subsequently, the data processing unit 40 performs an inner product projection operation between the temporal vector corresponding to the full field of view pixels and the secondary targeting basis to obtain the final targeting weight matrix after secondary correction.

[0037] S70, the data processing unit 40 performs numerical mapping on the final target weight matrix, converts it into a two-dimensional enhanced grayscale image, and performs connected component analysis on the two-dimensional enhanced grayscale image to mark the coordinates of continuous patches and output the corresponding AgInS quantum dot target recognition results.

[0038] The following detailed description of the processing steps in the technical solution of the present invention will further illustrate the process.

[0039] See attached document Figure 3 The following description, in conjunction with this embodiment, further illustrates step S10, which may specifically include the following steps: S101, the timing controller 30 generates periodic level trigger signals and sends them to the excitation light source 10 and the gated camera 20 respectively.

[0040] Upon receiving the trigger signal, the excitation source 10 emits a narrow-pulse excitation light towards the target region, ensuring that the pulse width is less than the fluorescence decay lifetime of the target quantum dot and biological tissue. This reduces the overlapping effect of the excitation light on the observation of fluorescence time-series evolution. During the return light acquisition process, the excitation band component is suppressed, and the signal corresponding to the emission band of the AgInS quantum dot is selectively acquired to reduce the interference of reflected and scattered excitation light on subsequent lifetime time-series reconstruction.

[0041] The timing controller 30 synchronously sends a shutter delay opening command to the gate camera 20, and sets the total number of delay time windows within a single acquisition cycle to [value missing]. . The excitation wavelength can be preset according to the fluorescence lifetime range of the target quantum dot, typically ranging from tens to hundreds of windows. In one embodiment, the excitation center wavelength can be 405 nm, the pulse width can be 100 ps, ​​and the repetition frequency can be 1 MHz. Option 64, delay time The starting point can be 0 ns, the step size can be 5 ns, and the ending point can be 315 ns. The integration time is... 5ns is acceptable.

[0042] For the A delay time window, of which The timing controller 30 controls the gated camera 20 after the excitation light pulse ends. Keep the shutter open at all times and during the continuous integration time. Then close the shutter. (Following the window number) Increment, delay time The time increments accordingly. The settings can be adjusted based on the photosensitivity and emission signal intensity of the acquisition device. Using the aforementioned delayed step-by-step acquisition method, the continuous fluorescence decay process can be discretized into multiple time-sampling slices, allowing for the reconstruction of fluorescence dynamics characteristics based on the sampling intensity at different time points.

[0043] The internal optical oscillation structure of the excitation light source 10 and the electronic shutter drive circuit of the gated camera 20 can be implemented using existing commercial devices or mature technical solutions, and the specific structures will not be elaborated here.

[0044] S102, the data processing unit 40 receives discrete two-dimensional image frames output by the gated camera 20 in each delay time window via a high-speed data bus.

[0045] During data reception, the data processing unit 40 verifies the number of received frames. If the number of received sequence frames is less than the preset total... If a frame is lost during the current acquisition cycle, the data processing unit 40 discards the data for that acquisition cycle and sends a re-acquisition command to the timing controller 30 via the communication link between the two units to avoid affecting subsequent matrix operations due to missing data.

[0046] In one embodiment, if the number of received frames still does not reach the target after three consecutive resampling attempts... If the current period is invalid, the data processing unit 40 will mark the current period as an invalid acquisition period and output an error flag, and then re-enter the processing flow in the next acquisition period.

[0047] Under normal reception conditions, the physical pixel array width of each frame of a two-dimensional image is denoted as: The height is recorded as The pixel grayscale values ​​in each frame of the two-dimensional image represent the local micro-regions of the target area at the corresponding integration time. The number of photons captured cumulatively within the cell.

[0048] In one embodiment, the data processing unit 40 constructs a three-dimensional spatiotemporal data matrix. Before or after, dark field subtraction, baseline correction, abnormal pixel correction, saturated pixel removal and / or gain uniformization can be performed on the two-dimensional image frames under each delay time window to improve the consistency between online acquired data and offline calibration reference vector in the data domain.

[0049] S103, the data processing unit 40 processes the data according to the receiving order of the two-dimensional image frames. Frame discrete image data are stacked and aligned along the time dimension.

[0050] Through the stacking operations described above, a continuous multidimensional array index is established for each pixel position in the storage space, thereby forming a three-dimensional spatiotemporal data matrix that carries luminous intensity information. Three-dimensional spatiotemporal data matrix The dimension can be represented as .

[0051] S104, Data processing unit 40 in the three-dimensional spatiotemporal data matrix A mapping relationship between spatial coordinates and the time axis is established to extract the temporal features corresponding to each spatial pixel.

[0052] Let the two-dimensional pixel coordinates be ,in , For any spatial coordinates The data processing unit 40 reads the data at that location sequentially along the time dimension. The corresponding gray values ​​under each delay time window are arranged to form a one-dimensional time-series signal column vector. Its mathematical model can be expressed as: ; in, Representing spatial coordinates The fluorescence temporal decay feature vector corresponding to the pixel has a dimension of ; This indicates that the cell is in the th order. Each time delay window corresponds to the luminous intensity value in the two-dimensional image; superscript This indicates transpose, used to convert a row vector into a column vector to accommodate subsequent linear algebra operations.

[0053] Using the extraction method described above, the fluorescence decay curve of a single pixel over time can be converted into a temporal feature vector in a high-dimensional algebraic space, thus providing a data foundation for subsequent interference separation operations based on linear orthogonal projection.

[0054] See attached document Figure 4 The following description, in conjunction with this embodiment, further illustrates step S20, which may include the following steps: S201, the data processing unit 40 accesses the non-volatile storage medium within the system and loads the pre-calibrated reference data offline.

[0055] In this embodiment, the reference data includes the AgInS quantum dot long-lived standard time decay vector used to characterize the target signal under test. And the short-lived initial prior vector of in vivo endogenous autofluorescence used to characterize background interference signals. Both reference vectors mentioned above can be calibrated under pure in vitro sample conditions using the delay time window parameters corresponding to the aforementioned time-synchronous acquisition mechanism. Preferably, the offline calibration process and the online acquisition process are similar in terms of excitation light center wavelength, optical filtering conditions, and gating integration time. Maintain consistency in key parameters such as delay time window settings, camera gain, and optical path configuration. If there are differences between offline calibration parameters and online acquisition parameters, recalibrate, resample the time axis, perform interpolation alignment, and / or gain mapping on the offline calibration vector before using it for subsequent calculations.

[0056] in, It can be obtained by collecting in vitro standard samples containing only AgInS quantum dots. The calibration vector can be obtained from live background samples or blank tissue samples that do not contain AgInS quantum dots. For each type of calibration sample, it is preferable to continuously collect at least 20 complete cycles, and perform dark noise subtraction, element-wise averaging and L2 normalization on the time-series vector corresponding to each cycle in sequence to generate the corresponding calibration vector, which is then stored in a non-volatile storage medium.

[0057] The data processing unit 40 will load the obtained and Initialize as column vectors corresponding to the temporal features, with all of them having an algebraic dimension of 1. .

[0058] S202, Data processing unit 40 sets initial prior vectors for short-lived endogenous autofluorescence in vivo. Perform L2 norm calculation to generate normalized background basis vectors. .

[0059] The normalization process can be represented as: ; in, Represents column vectors The L2 norm is calculated. During the calculation process, the data processing unit 40 performs a minimum threshold check on this norm; if the calculation result is lower than a preset machine precision lower limit, the loaded initial prior vector is deemed invalid, and the current calculation process is terminated to avoid anomalies where the denominator is zero. The machine precision lower limit can be set according to the data representation precision of the computing hardware, for example, it can be 10. -6 Up to 10 -8 Magnitude.

[0060] After the above normalization process, the background basis vectors The directional characteristics of the time-series decay pattern of endogenous autofluorescence in living organisms are preserved, thereby reducing the impact of absolute light intensity fluctuations and differences in acquisition scale on subsequent calculations.

[0061] S203, the data processing unit 40 uses the Gram-Schmidt orthogonalization method to process the long-lived standard time decay vector of the AgInS quantum dot. Perform projection stripping operation to extract the basis vectors relative to the background. Orthogonal residual components.

[0062] Specifically, the data processing unit 40 first calculates exist The projection components in the direction, and then... Subtracting the projected components yields the orthogonal residual vector. It can be represented as: ; in, The result of the vector inner product is used to characterize the linear correlation component between the target signal and the background luminescence features; This represents the projection vector composed of the background component. Through the above projection stripping, it is possible to... The component orthogonal to the background basis direction is separated to characterize the temporal characteristics of the target quantum dot after suppression by common background features.

[0063] S204, Data Processing Unit 40, Orthogonal Residual Vectors Perform L2 norm normalization to generate an initial targeting basis for subsequent temporal feature separation. .

[0064] The normalization process can be represented as: ; in, Represents orthogonal residual vectors The L2 norm. Normalized It serves as the initial targeting basis for subsequent pixel temporal vector projection calculations.

[0065] Under in vivo optical imaging conditions, the fluorescence lifetime of AgInS quantum dots is typically on the order of hundreds of nanoseconds, while the autofluorescence lifetime of endogenous tissues is typically on the order of single-digit nanoseconds. Based on these temporal differences, a long-lifetime standard time decay vector... With short-lifetime initial prior vector They typically exhibit significant linear independence, thus making the orthogonal residual vectors It maintains a non-zero norm under normal acquisition conditions. This provides a stable mathematical basis for the construction of the initial target substrate.

[0066] See attached document Figure 5 The following description, in conjunction with this embodiment, further illustrates step S30, which may include the following steps: S301, Data processing unit 40 acquires three-dimensional spatiotemporal data matrix Spatial resolution and And based on this, a corresponding two-dimensional spatial coordinate system is established.

[0067] In a two-dimensional spatial coordinate system, the data processing unit 40 traverses all spatial pixels within the entire field of view and, based on the currently accessed spatial coordinates... From the three-dimensional spatiotemporal data matrix Extract the corresponding one-dimensional time-series signal column vector .

[0068] In some implementations, to improve the processing efficiency of full-field pixel traversal and subsequent operations, the data processing unit 40 can perform parallel processing of vector extraction and inner product operations corresponding to different spatial coordinates based on a parallel computing architecture. The relevant parallel scheduling methods can be implemented using existing general-purpose computing programming models, which will not be elaborated here.

[0069] S302, the data processing unit 40 processes the extracted time-series signal column vector. Call the initial target substrate generated in step S20 .

[0070] Data processing unit 40 converts the timing signal column vector Projected onto the initial target substrate The spatial coordinates are obtained by calculating the direction of the vector and using the dot product operation. The initial target weight scalar at the location. The operation can be expressed as: ; in, Representing spatial coordinates The weight scalar obtained after preliminary projection, Represents a column vector of time-series signals The transpose. Due to the initial target substrate. The short-lived autofluorescence components have been orthogonally stripped; therefore, the time-series signal column vector... and The inner product result can be used to characterize the signal component in the current pixel that matches the luminescence characteristics of the AgInS quantum dot.

[0071] S303, the data processing unit 40 processes the weight scalar obtained from the initial projection. Perform non-negative truncation.

[0072] During numerical calculations, the data is affected by factors such as acquisition noise, background fluctuations, and deviations between the reference vector and the actual state. Negative values ​​may occur. Given that optical luminescence intensity should not be negative in a physical sense, the data processing unit 40 applies boundary constraints to the initial projection results: when... When it is less than 0, assign it the value 0; when When the value is greater than or equal to 0, its original value is retained. After processing, the spatial coordinates are obtained. Initial target weight value at the location .

[0073] The nonnegative truncation process can be represented as: ; The above processing can filter out negative values ​​that have no physical meaning, thus preventing abnormal values ​​from being passed on to subsequent processing stages.

[0074] S304, Data processing unit 40 completes the three-dimensional spatiotemporal data matrix All spatial coordinates After traversing and calculating, the initial target weight values ​​corresponding to each spatial location are... Arrange the targets in two dimensions according to their original spatial coordinates to generate the initial target weight matrix. .

[0075] Initial target weight matrix The spatial dimension is This process maps the temporal features in three-dimensional spatiotemporal data into a two-dimensional spatial distribution result, thereby obtaining a weight matrix that characterizes the initial spatial distribution state of the AgInS quantum dot signal, which serves as input for subsequent background segmentation and morphological processing.

[0076] See attached document Figure 6 The following description, in conjunction with this embodiment, further illustrates step S40, which may include the following steps: S401, Data processing unit 40 reads the initial target weight matrix The Otsu method was used to determine the global adaptive segmentation threshold. .

[0077] Specifically, the data processing unit 40 pairs the initial target weight matrix The weight scalars in the image are statistically analyzed, and the segmentation threshold is determined based on the inter-class variance of the foreground and background pixels. Otsu's method is based on the statistical characteristic that the target and background are separable in the image histogram. It searches for the gray level that maximizes the inter-class variance to obtain a global threshold for distinguishing the target signal region from the background region. .

[0078] To avoid threshold failure under extreme imaging conditions, the data processing unit 40 compares the maximum inter-class variance with a preset noise floor variance threshold after obtaining it. When the maximum inter-class variance is lower than the noise floor variance threshold, it is determined that the current field of view lacks effective contrast, the Otsu method branch is terminated, and a fixed empirical threshold calibrated offline is used as the global adaptive segmentation threshold. .

[0079] In one embodiment, the noise floor variance threshold can be 1.2 to 1.5 times the mean variance between classes of blank background samples, for example, 1.3 times; the fixed empirical threshold can be the initial target weight matrix. 15% to 25% of the maximum value, for example, take 20%.

[0080] The iterative solution process of Otsu's method and the search for maximizing the inter-class variance can be implemented by existing digital image analysis algorithms, which will not be elaborated here.

[0081] S402, the data processing unit 40 determines the segmentation threshold... Generate a binarized target region mask .

[0082] Specifically, the data processing unit 40 constructs the initial target weight matrix. A two-dimensional logic array of uniform size is used, and the weight value of each spatial position is determined sequentially. When Greater than or equal to At that time, Assign a value of 1; when Less than At that time, The value is assigned to 0. Its discriminant relation can be expressed as: ; in, Representing spatial coordinates Mask value at that location, This represents the initial targeting weight at the corresponding location. After the above processing, the region in the mask with a value of 1 corresponds to the suspected AgInS quantum dot luminescent region.

[0083] In generation Subsequently, the data processing unit 40 further calculates the proportion of pixels with a value of 1 in the mask. When the proportion exceeds a preset upper limit threshold, it is determined that the current field of view is approximately full target coverage. In this cycle, the dynamic background resampling in step S50 is not performed, and in step S60, the short-lived initial prior vector of in vivo endogenous autofluorescence loaded in step S20 continues to be used. Participate in secondary targeted matrix generation.

[0084] When the proportion is lower than the preset lower limit threshold, the current field of view is determined to be approximately the entire background. The data processing unit 40 marks the current period as a low signal period and terminates steps S50 and S60 of this period. In the subsequent step S70, the empty recognition result is directly output.

[0085] In one embodiment, the upper threshold can be 95%, and the lower threshold can be 0.5%.

[0086] S403, Data Processing Unit 40 for Target Region Mask Perform a Boolean negation to obtain the initial background region.

[0087] After inversion, the pixel value corresponding to the target area in the original mask becomes 0, and the pixel value corresponding to the original background area becomes 1, thereby converting the object to be processed from the suspected target area into the suspected background area.

[0088] S404, the data processing unit 40 performs morphological erosion processing on the inverted preliminary background area to generate a high-confidence binary background mask. .

[0089] In live-cell fluorescence imaging, signal aliasing often occurs at the boundary between the target and background regions. Directly using these pixels as background references can easily introduce long-lived fluorescence components from the target into the background modeling process. To reduce the impact of boundary aliasing pixels on subsequent dynamic background substrate extraction, the data processing unit 40 loads pre-defined morphological structural elements. The erosion operation is performed on the initial background area, causing the background area to shrink inward in space, thereby eliminating pixels close to the target boundary.

[0090] The processing relationship can be represented as: ; in, This serves as the final high-confidence background binary mask. This represents the Boolean negation operation on a two-dimensional matrix; This refers to the erosion operator in mathematical morphology. This represents a defined two-dimensional morphological structural element.

[0091] In practical implementation, structural elements A cross-shaped or rectangular window can be used, and its size can be set to an odd number of dimensions such as 3×3 or 5×5. The specific parameters can be determined according to the spatial resolution and optical diffraction characteristics of the imaging system.

[0092] To avoid excessive erosion of the background area due to overly large structuring elements, the data processing unit 40 performs statistics after each erosion. The number of pixels with a value of 1. When the total number of non-zero pixels falls below a preset lower safety threshold, the structuring element is automatically reduced. The erosion operation is then re-executed to retain a sufficient number of background samples.

[0093] In one embodiment, structural element The minimum size is limited to 3×3; when the structuring element ε has been reduced to the minimum size and If the number of pixels with a value of 1 is still lower than the lower safety threshold, the current dynamic background resampling process is terminated, and the initial prior vector of short-lived endogenous autofluorescence loaded in step S20 is used in subsequent calculations in step S60.

[0094] High-confidence background binary mask after etching treatment Pixels with a value of 1 are located inside the background far from the target region boundary and can be used as high-confidence background samples for subsequent dynamic background basis extraction. This process reduces the impact of target boundary crosstalk on the background modeling results.

[0095] See attached document Figure 7 The following description, in conjunction with this embodiment, further illustrates step S50, which may include the following steps: S501, the data processing unit 40 establishes the inverse mapping relationship between the two-dimensional morphological mask and the three-dimensional spatiotemporal data matrix.

[0096] Specifically, the data processing unit 40 reads the high-confidence background binary mask output in step S40. and combine it with the three-dimensional spatiotemporal data matrix constructed in step S10. Perform spatial coordinate alignment. For any spatial coordinate in the 2D mask... In terms of indexing, it corresponds to a three-dimensional spatiotemporal data matrix. The time axis data columns at the same horizontal and vertical coordinate positions are used. Based on this mapping relationship, the corresponding position of the background region in the original three-dimensional spatiotemporal data can be determined.

[0097] S502, before performing background feature extraction, the data processing unit 40 processes the high-confidence background binary mask. The number of valid background pixels in the data is checked.

[0098] The data processing unit iterates through the high-confidence background binary mask 40 times. Count the total number of pixels with an attribute value of 1, and record this as the number of valid background pixels. To avoid the mean calculation failing or division by zero anomalies due to insufficient background sample size, the data processing unit 40 will... Compared with the preset safety lower threshold Compare the safety lower limit thresholds. It can be set to a certain percentage of the total number of pixels in the entire field of view, usually between 0.1% and 0.5% of the total number of pixels. The specific value can be set according to the system signal-to-noise ratio and calculation accuracy.

[0099] when When the number of samples in the current high-confidence background region is insufficient, the data processing unit 40 terminates the current dynamic background substrate resampling process and continues to use the short-lived in vivo endogenous autofluorescence initial prior vector loaded in step S20 for subsequent calculations. When the current number of background samples meets the extraction conditions, the data processing unit 40 proceeds to the subsequent background feature calculation step.

[0100] S503, when the number of background samples meets the requirements, the data processing unit 40 extracts the temporal vector corresponding to each pixel in the high-confidence background region according to the inverse mapping relationship, and generates a dynamic background basis vector. .

[0101] Specifically, the data processing unit 40 addresses all data that satisfy the following conditions in a two-dimensional plane. Spatial coordinates, and from the three-dimensional spatiotemporal data matrix Extract the one-dimensional time-series signal column vector corresponding to each spatial coordinate. Subsequently, the data processing unit 40 accumulates the extracted time-series signal column vectors element by element along the time dimension, and divides the accumulation result corresponding to each time node by the number of effective background pixels. To obtain the dynamic background basis vectors .

[0102] In one embodiment, dynamic background basis vectors The mean value can be obtained directly from the unnormalized original temporal vector in the high-confidence background region to preserve the actual energy distribution characteristics of the background interference under the current field of view. In another embodiment, the temporal vector of each background pixel can be normalized first, and then the mean value operation can be performed to highlight the background attenuation morphology characteristics.

[0103] Dynamic background basis vectors The calculation process can be expressed as: ; in, This represents the dynamic background basis vector, with dimension . ; This indicates a high-confidence background binary mask. Summation is performed on the temporal vectors corresponding to all spatial coordinates with an attribute value of 1. Representing spatial coordinates The time-series signal column vector at the location.

[0104] Through the above processing, the background reference features can be updated using the actual sampling results of the background region within the current field of view, thereby making the background substrate used in subsequent calculations closer to the background interference state under the current imaging conditions. At the same time, averaging the multi-pixel temporal vectors can also reduce the impact of single-point acquisition noise on the stability of the background reference vector to some extent.

[0105] See attached document Figure 8 The following description, in conjunction with this embodiment, further illustrates step S60, which may include the following steps: S601, Data processing unit 40 performs background reference vector update.

[0106] Specifically, the data processing unit 40 receives the dynamic background basis vector output in step S50. and in system memory as dynamic background basis vectors The short-lived initial prior vector of in vivo endogenous autofluorescence, which was loaded in the initial stage, is replaced. Through the above update, the actual temporal characteristics of the background region under the current field of view can be introduced into the subsequent calculation process, so as to reduce the impact of the deviation between offline static prior and real-time imaging state on the background stripping result.

[0107] S602, Data processing unit 40 is based on dynamic background basis vectors Generate secondary target substrate .

[0108] First, the data processing unit 40 pairs dynamic background basis vectors Performing L2 norm normalization yields a normalized dynamic background vector. The calculation process can be expressed as follows: ; in, Represents dynamic background basis vectors The L2 norm. Data processing unit 40 performs a minimum threshold check on the norm to avoid division by zero anomalies. The minimum safety threshold can be set according to the data representation precision of the computing hardware, for example, it can be 10. -8 Magnitude. When When the value is below the minimum safety threshold, the data processing unit 40 does not use... The background reference is updated, while the initial short-lived in vivo endogenous autofluorescence vector loaded in step S20 is used in subsequent calculations.

[0109] After the verification is passed, the data processing unit 40 uses the Gram-Schmidt orthogonalization method to calculate the long-lifetime standard time decay vector. In the normalized dynamic background vector The projection components in the direction are used to obtain the orthogonal residual vector. The calculation process can be expressed as follows: ; in, Represents the long-lifetime standard time decay vector With the normalized dynamic background vector The inner product result.

[0110] Obtaining orthogonal residual vectors Subsequently, the data processing unit 40 further verifies its L2 norm. When the orthogonal residual vector... When the L2 norm falls below the minimum safety threshold, the data processing unit 40 terminates the current secondary target basis update and reverts to using the initial target basis. Participate in subsequent calculations; when orthogonal residual vectors When the L2 norm is not lower than the minimum safety threshold, the data processing unit 40 performs normalization processing on it to obtain the secondary target substrate. The calculation process can be expressed as follows: ; in, Represents orthogonal residual vectors The L2 norm. The secondary targeting substrate obtained after the above processing. Used to characterize the target temporal features after orthogonal stripping relative to the current dynamic background features.

[0111] S603, Data Processing Unit 40 utilizes a secondary targeting substrate Perform corrected projection operations on all pixels in the field of view.

[0112] Specifically, the data processing unit 40 re-traverses the three-dimensional spatiotemporal data matrix. Spatial coordinates in And extract the corresponding one-dimensional time-series signal column vector. Subsequently, the data processing unit 40 calculates the timing signal column vector. With secondary target base The inner product between them is used to obtain the current spatial coordinates. The second projection weight scalar at the given location. Its calculation process can be expressed as: ; in, Representing spatial coordinates The weights of the second projection without boundary constraints. Represents a column vector of time-series signals The transpose of the signal. Through the above projection calculation, the original mixed signal can be separated again based on the updated background reference features to further suppress residual background interference.

[0113] S604, the data processing unit 40 performs non-negativity constraint processing on the secondary projection weight scalar and generates the final target weight matrix. .

[0114] Considering the influence of acquisition noise and model errors, the secondary projection results may contain negative values, which do not have any actual significance in terms of luminous intensity. The data processing unit has 40 pairs of... Enforce boundary constraints: when When it is less than zero, set it to zero; when When the value is greater than or equal to zero, its original value is retained. This yields the final target weight value for a single point. It can be represented as: ; in, This represents the effective signal weight scalar after non-negativity constraints.

[0115] After completing the traversal calculation of the entire field of view spatial coordinates, the data processing unit 40 processes all data according to the corresponding positions of each spatial coordinate. Perform two-dimensional array arrangement to generate the final target weight matrix. Final target weight matrix The spatial dimension is .

[0116] Final target weight matrix This is used to characterize the intensity distribution of AgInS quantum dot signals in two-dimensional space after dynamic background updating and quadratic orthogonal stripping. In some embodiments, the final targeting weight matrix... It can be directly output to a display terminal for observation, or transmitted to a downstream medical image processing module for further analysis.

[0117] See attached document Figure 9 The following description, in conjunction with this embodiment, further illustrates step S70, which may include the following steps: S701, Data processing unit 40 reads the final target weight matrix obtained from the calculation. The floating-point weight values ​​are then mapped to the grayscale range corresponding to the preset bit depth to generate an enhanced grayscale image.

[0118] Specifically, the final target weight matrix The elements in the data processing unit are typically floating-point scalars. To adapt to the data interface of the display terminal or subsequent image processing module, the data processing unit 40 extracts the final targeting weight matrix. Maximum weight value in and minimum weight value And based on the maximum weight value and minimum weight value Establish a linear stretching mapping relationship. The mapping process can be represented as: ; in, Representing spatial coordinates The mapped grayscale value, where D represents the preset image bit depth parameter. When using 8-bit display mode, D can be 8; when using 16-bit display mode, D can be 16.

[0119] To avoid the anomaly of a denominator of zero under conditions of no valid signal, the data processing unit 40 performs a certain step before performing the linear stretching mapping. and Threshold validation is performed on the difference between them. and When the difference is lower than the preset floating-point precision lower limit, it is determined that there is no effective target signal in the current field of view, and all pixel values ​​of the enhanced grayscale image are set to zero. At the same time, the current period is marked as a period without effective target signal, and the empty recognition result is output in steps S702 and S703. The floating-point precision lower limit can be set according to the system hardware and software configuration, for example, it can be 10. -6 Up to 10 -8 Magnitude.

[0120] After the above processing, the final target weight matrix can be obtained while preserving the relative intensity distribution relationship of the target signal. Convert to an enhanced grayscale image suitable for display and subsequent analysis.

[0121] S702, the data processing unit 40 performs connected component analysis on the enhanced grayscale image to extract the spatial range of the target luminescent patch.

[0122] Specifically, the data processing unit 40 segments the enhanced grayscale image based on a preset activation threshold, filters out active pixels that meet the threshold conditions, and merges connected regions of the active pixels according to the adjacency relationship between pixels in a two-dimensional spatial coordinate system. In one embodiment, the activation threshold can be 20% of the maximum grayscale value of the enhanced grayscale image, and the connected region analysis can use an eight-connectivity labeling algorithm.

[0123] After completing the initial region extraction, the data processing unit 40 further performs area statistics on each region to be determined and compares the statistical results with a preset lower area threshold. The lower area threshold can be preset according to the minimum expected size of the target and the optical magnification of the system, for example, it can be set to 5 to 20 pixels. For connected regions with an area smaller than the lower area threshold, the data processing unit 40 determines them as isolated noise and removes them; for connected regions with an area not smaller than the lower area threshold, they are retained as effective target regions.

[0124] Through the above processing, the effective connected region corresponding to the AgInS quantum dot targeting signal can be separated from the enhanced grayscale image.

[0125] S703, the data processing unit 40 calculates the geometric parameters of the selected effective connected regions and generates the corresponding recognition results.

[0126] For any valid connected region, the data processing unit 40 calculates its geometric centroid coordinates based on the spatial coordinates of all valid pixels within that region. The calculation process can be represented as follows: ; ; in, This indicates the total number of pixels contained in the currently active connected region. and They represent the first The x and y coordinates of each valid pixel.

[0127] While calculating the geometric centroid coordinates, the data processing unit 40 can also extract the outer contour boundary of the effective connected region and generate contour parameters to characterize the target shape. The contour parameters may include minimum bounding rectangle parameters or closed polygon parameters.

[0128] Finally, the data processing unit 40 outputs the enhanced grayscale image, geometric centroid coordinates, and contour parameters as structured data to the display interface or downstream medical processing workstation via the system's external communication interface.

[0129] Through the above steps, the final target weight matrix can be obtained. The results are converted into visual recognition results and output spatial positioning information and morphological parameters of the target area.

[0130] To further assist those skilled in the art in understanding the practical application effects and technical advantages of the present invention, a specific application example is provided below, using a specific AgInS quantum dot mouse in vivo tumor-targeting imaging experiment.

[0131] In this embodiment, the hardware parameters of the image enhancement and recognition system are specifically configured as follows: the excitation light source 10 uses a picosecond pulsed semiconductor laser with a center wavelength of 405nm, a pulse width of 100ps, and a pulse repetition frequency of 1MHz; the gated camera 20 uses a picosecond-level gated ICCD camera with an image intensifier, and its two-dimensional spatial resolution is configured as 256×256 pixels. The delayed acquisition sequence parameter controlled by the timing controller 30 is set to the total number of delay time windows. The value is 64, the initial delay time is 0ns, the step size is 5ns, and the integration time is... Fixed at 5ns.

[0132] In the live sample preparation stage, BALB / c nude mice were selected as the animal model, and AgInS quantum dot probes modified with targeting were injected via the tail vein. Three test sample groups were established to verify the robustness of the algorithm: Sample A was set as a subcutaneous superficial tumor model with relatively simple background interference; Sample B was set as an in situ liver tumor model, accompanied by extremely strong and complex blood supply autofluorescence background interference; Sample C was set as a blank control group that received an equal volume of physiological saline.

[0133] Before the system is online, the data processing unit 40 uses pure AgInS quantum dot aqueous solution and normal nude mouse ex vivo tissue without probe injection to extract long-lived standard time decay vector and in vivo endogenous autofluorescence short-lived initial prior vector under the same time parameters, and saves them to non-volatile storage medium.

[0134] During the dynamic operation and reconstruction phase of the algorithm, the data processing unit 40 receives the acquired sequence and constructs a three-dimensional spatiotemporal data matrix with dimensions of 256×256×64. When performing the morphological processing step, the global adaptive segmentation threshold is calculated using the Otsu method. For the in situ liver tumor model of sample B, due to changes in the pH of the perivascular tissue fluid and blood perfusion, the local autofluorescence lifetime distribution exhibits approximately a 15% feature drift compared to the offline prior model. The data processing unit 40 accurately extracts the dynamic background basis vector containing this transient attenuation distortion by executing a coordinate inverse mapping mechanism, and then uses this updated vector to reconstruct the secondary targeting basis, correcting the aforementioned 15% drift error from the algorithm's underlying layer.

[0135] See attached document Figure 10 This figure shows a comparison of the target identification results of multiple live tumor samples at different processing stages in this embodiment. The bright field information of the background image has been pre-processed by background stripping and smoothing to highlight the target area observation effect. As shown in the figure, in the CW steady-state fluorescence mode, the luminescence of AgInS quantum dots in samples A and B is severely mixed with the extremely strong autofluorescence of the digestive tract and liver, making it impossible to accurately locate the centroid of the target area. In the static prior orthogonal projection mode, due to the prior feature drift, a large number of unstripped vascular autofluorescence artifacts remain at the edge of the tumor in sample B. However, in the dynamic background correction projection mode provided by this invention, after dynamic background resampling and secondary projection correction, the background noise of samples A and B is completely suppressed, the high-intensity autofluorescence around the liver and blood vessels is accurately stripped, the true outline of the target plaque is clearly presented, the centroid coordinate calculation accuracy reaches the sub-millimeter level, and the output of the blank control group sample C is a pure dark field, without triggering false positives.

[0136] See attached document Figure 11 The paper presents a quantitative comparison of the signal-to-background ratio (SNR) under different algorithms. Based on the quantitative statistical results, under the complex in vivo microenvironment of sample B, the average SNR of the CW steady-state fluorescence mode is only 1.35, the average SNR based on the static prior orthogonal projection method is 3.82, while the average SNR of the dynamic background-corrected projection output provided by the present invention jumps to 8.74. Compared with traditional static orthogonal projection, the present invention achieves a significant improvement in the contrast of the targeted signal, greatly suppressing physiological background noise.

[0137] See attached document Figure 12The table shows the subject operating characteristic (SOC) curve for targeted classification. Cross-validation results of the classifier based on the weight matrix data derived from the system show that the area under the curve (AUC) of the proposed scheme reaches 0.985 when distinguishing between probe-containing tissue and tissue with strong background interference, significantly better than the 0.871 of the static prior algorithm. These multi-dimensional validation data fully demonstrate that the dynamic background-based resampling mechanism based on morphological mask inverse mapping and the quadratic orthogonalization correction logic in this invention possess extremely high environmental adaptability, robustness, and targeted recognition accuracy in the complex microenvironment of living organisms.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for enhancing and recognizing bio-imaging images using AgInS quantum dots, characterized in that, Includes the following steps: Two-dimensional image frames acquired under different delay time windows are received, and the two-dimensional image frames are stacked sequentially along the time dimension to construct a three-dimensional spatiotemporal data matrix; Load the pre-calibrated AgInS quantum dot long-lived standard time decay vector and the in vivo endogenous autofluorescence short-lived initial prior vector; perform orthogonalization operation on the long-lived standard time decay vector and the short-lived initial prior vector to generate the initial target substrate; Extract the temporal vector corresponding to each spatial pixel in the three-dimensional spatiotemporal data matrix, and calculate the inner product between the temporal vector and the initial targeting basis to generate the initial targeting weight matrix; The initial target weight matrix is ​​subjected to global adaptive segmentation and morphological processing based on erosion operation to extract a binarized high-confidence background mask; The high-confidence background mask is reverse-mapped to the three-dimensional spatiotemporal data matrix, the temporal vector corresponding to the mask coordinates is extracted, and the arithmetic mean is performed to generate a dynamic background basis vector characterizing the current physiological interference features. The orthogonalization operation is re-performed by replacing the short-lifetime initial prior vector with the dynamic background basis vector to generate a secondary targeting basis; Perform an inner product projection operation between the temporal vector corresponding to the full field of view pixels and the secondary targeting basis to obtain the final targeting weight matrix; The final target weight matrix is ​​numerically mapped and converted into a two-dimensional enhanced grayscale image. Connectivity analysis is then performed on the two-dimensional enhanced grayscale image to output the recognition result.

2. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of performing an orthogonalization operation on the long-lifetime standard time decay vector and the short-lifetime initial prior vector to generate an initial target basis includes: Perform L2 norm calculation on the short-lifetime initial prior vector to generate a normalized background basis vector; The Gram-Schmidt orthogonalization method is used to calculate the projection component of the long-lifetime standard time decay vector in the direction of the normalized background basis vector, so as to extract the orthogonal residual vector; The orthogonal residual vector is subjected to L2 norm normalization to generate the initial targeting basis.

3. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of calculating the inner product between the time-series vector and the initial target basis to generate the initial target weight matrix includes: Calculate the inner product between the temporal vector and the initial target basis to obtain the initial target weight scalar at the spatial coordinates; The initial target weight scalar is subjected to non-negative truncation. If the initial target weight scalar is less than zero, it is assigned the value of zero; if the initial target weight scalar is greater than or equal to zero, its original value is retained to obtain the initial target weight value. The initial target weight values ​​corresponding to each spatial location are arranged in two dimensions according to the original spatial coordinates to generate the initial target weight matrix.

4. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of performing global adaptive segmentation and morphological processing based on erosion operation on the initial target weight matrix to extract a binarized high-confidence background mask includes: The Otsu method is used to determine the global adaptive segmentation threshold based on the inter-class variance of foreground and background pixels in the initial targeting weight matrix. A binarized target region mask is generated based on the global adaptive segmentation threshold for the initial target weight matrix; Perform a Boolean inversion on the target region mask to obtain the preliminary background region; The high-confidence background mask is generated by performing an erosion operation on the preliminary background region using preset morphological structural elements.

5. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 4, characterized in that, After generating the target region mask, the process further includes: The percentage of pixels with a value of 1 in the target region mask is counted. If the percentage of pixels is higher than a preset upper limit threshold, the step of generating the dynamic background basis vector is not executed, and the short-lived initial prior vector is used to participate in the generation of the secondary target basis. If the pixel percentage is lower than the preset lower threshold, then an empty recognition result is output directly.

6. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of inversely mapping the high-confidence background mask to the three-dimensional spatiotemporal data matrix, extracting the temporal vector corresponding to the mask coordinates, and performing an arithmetic mean operation to generate a dynamic background basis vector characterizing the current physiological interference features includes: Count the number of valid background pixels in the high-confidence background mask; When the number of effective background pixels is not less than the preset safety lower limit threshold, spatial coordinates that satisfy the mask value of 1 are addressed in the two-dimensional plane, and the temporal vectors corresponding to each spatial coordinate are extracted from the three-dimensional spatiotemporal data matrix. The extracted time-series vectors are accumulated element by element along the time dimension, and the accumulated result corresponding to each time node is divided by the number of effective background pixels to obtain the dynamic background basis vector.

7. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of replacing the short-lifetime initial prior vector with the dynamic background basis vector and re-performing orthogonalization to generate a secondary targeting basis includes: The dynamic background basis vectors are normalized using the L2 norm to obtain normalized dynamic background vectors; The projection component of the long-lifetime standard time decay vector in the direction of the normalized dynamic background vector is calculated using the Gram-Schmidt orthogonalization method, and the orthogonal residual vector is obtained. The orthogonal residual vector is subjected to L2 norm normalization to obtain the secondary targeting basis.

8. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The step of performing an inner product projection operation between the temporal vector corresponding to the full field of view pixels and the secondary targeting basis to obtain the final targeting weight matrix includes: Calculate the inner product between the temporal vector corresponding to each spatial coordinate and the secondary target basis to obtain the secondary projection weight scalar; Boundary constraint processing is performed on the secondary projection weight scalar. When the secondary projection weight scalar is less than zero, it is set to zero; when the secondary projection weight scalar is greater than or equal to zero, its original value is retained to obtain the final target weight value. All the final target weight values ​​are arranged in a two-dimensional array according to their corresponding positions in spatial coordinates to generate the final target weight matrix.

9. The AgInS quantum dot bioimaging image enhancement and recognition method according to claim 1, characterized in that, The final target weight matrix is ​​numerically mapped to convert it into a two-dimensional enhanced grayscale image, and connected component analysis is performed on the two-dimensional enhanced grayscale image to output the recognition result, including: Extract the maximum and minimum weight values ​​from the final target weight matrix, establish a linear stretching mapping relationship based on the maximum and minimum weight values, map the floating-point weight values ​​to the grayscale range corresponding to the preset bit depth, and generate the two-dimensional enhanced grayscale image. The two-dimensional enhanced grayscale image is segmented based on a preset activation threshold, and active pixels that meet the threshold conditions are selected. The active pixels are then merged into connected regions based on their adjacency relationships. The area of ​​the merged connected regions is statistically analyzed, and connected regions with an area less than the preset lower area threshold are removed. Connected regions with an area not less than the preset lower area threshold are retained as valid connected regions. The geometric centroid coordinates are calculated based on the spatial coordinates of all valid pixels within the effective connected region, and the outer contour boundary is extracted to generate contour parameters, and the recognition result is output.

10. An AgInS quantum dot bioimaging image enhancement and recognition system, characterized in that, include: Excitation light source, used to emit periodic excitation light pulses towards the target area of ​​a living organism; A gated camera is positioned on the return optical path of the excitation light source to receive the fluorescence signal generated by the target tissue after excitation, and to acquire images of the fluorescence signal to output a two-dimensional image frame. A timing controller is communicatively connected to the excitation light source and the gated camera, respectively, and is used to output a synchronous trigger signal to control the light pulse emission period of the excitation light source, and to control the gated camera to perform shutter opening and closing operations within different delay time windows after the excitation light pulse; The data processing unit is communicatively connected to the gated camera and the timing controller, respectively, and is used to execute the AgInS quantum dot bioimaging image enhancement and recognition method as described in any one of claims 1-9.

Images