A multi-scale differential evaluation method for microscope real image data

Abstract

This invention discloses a multi-scale difference evaluation method for real microscope image data, belonging to the field of volumetric electron microscopy image processing technology. The invention first enhances the intensity variation in the difference between adjacent slices using Laplacian filtering, thereby highlighting irregular noise signals. Subsequently, multi-scale difference filtering (such as the difference-of-Gaussian pyramid) is used to extract texture features at different spatial scales to reflect the continuity and fidelity of the structure. The registration quality can be evaluated by the comprehensive characteristics of texture contrast, noise distribution, and information entropy in the multi-scale image. Compared with existing microscope image registration quality evaluation methods, the multi-scale difference evaluation algorithm proposed in this invention significantly improves the objectivity, stability, and adaptability of the evaluation.

Description

Technical Field

[0001] This invention relates to the field of volumetric electron microscopy image processing technology, and more specifically to a multi-scale differential evaluation method for real microscope image data. Background Technology

[0002] In volumetric electron microscopy imaging data analysis, image registration is a crucial step in achieving 3D reconstruction and quantitative measurement. However, for real microscope image data, obtaining a completely accurate "Ground Truth" image is nearly impossible due to various factors such as sample preparation, optical distortion, slice non-uniformity, and noise interference. This makes the objective evaluation of registration algorithm performance a significant challenge.

[0003] Existing methods for evaluating the registration quality of microscope images mainly focus on two directions: automated analysis based on image statistical features and qualitative evaluation based on manual annotation or structural comparison. In terms of automated statistical analysis, one evaluation method based on the contrast of the gray-level co-occurrence matrix (GLCM) quantifies the overall smoothness of the image stack by calculating the co-occurrence relationship of gray-level distributions in adjacent slices, assuming that higher smoothness indicates better registration. This method is computationally simple and can be automated, but it primarily reflects intensity continuity or texture smoothness rather than geometric correspondence, thus failing to effectively reveal spatial misalignments caused by translation, rotation, or local deformation, thereby reducing the reliability of the evaluation. In the direction based on manual annotation, 3D registration accuracy is evaluated by manually tracing the neuronal skeleton. The core idea is to use manually annotated neuronal structures as a benchmark, reflecting registration performance by measuring the spatial consistency of neuronal paths before and after registration. This method has high spatial resolution in assessing local structural accuracy, but it relies heavily on manual intervention, is time-consuming, and highly subjective. When the data scale reaches thousands of slices, this method is difficult to generalize. An evaluation strategy based on manual anchor points is proposed. This method manually defines a series of identifiable anchor points between adjacent slices and calculates the angular residuals and distance offsets before and after registration to estimate the geometric consistency between slices. Although this method is more intuitive in computation than the previous one, it still has two main problems: first, the distribution of anchor points is limited by human experience and cannot cover the entire image space; second, the angular and positional residuals only reflect the accumulation of errors in local areas and are difficult to reflect the global trend of overall registration quality.

[0004] Existing methods for evaluating the registration quality of microscope images generally rely on manual annotation or indirect indicators based on local gray-level statistics, which suffer from insufficient objectivity and low automation. Especially in real microscope data, due to the lack of ground truth images, traditional evaluation methods based on annotation, skeleton tracking, or anchor point residuals are difficult to accurately reflect the overall registration quality. At the same time, in practical applications, microscope images are often accompanied by complex noise interference, nonlinear deformation, and background drift, resulting in registration errors with strong spatial non-uniformity and multi-scale characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-scale differential evaluation method for real microscope image data, so as to objectively, automatically and comprehensively evaluate the registration quality with multi-scale perception capabilities in the absence of ground-based reference, thereby overcoming the shortcomings of the prior art.

[0006] This invention constructs a multi-scale difference filtering model to extract multi-level texture and structural feature information from the intensity difference between adjacent slices. It calculates the difference images between adjacent slices and performs difference analysis in a multi-scale space to simultaneously evaluate noise suppression and structural preservation. The core idea is to use the difference features between adjacent slice images to reflect the registration quality. Good registration should present a smooth texture in the difference image consistent with the changes in biological tissue structure, while poor registration will introduce random high-frequency noise.

[0007] Therefore, this invention aims to evaluate the registration effect from two complementary dimensions: (1) Noise Suppression – By analyzing the high-frequency residual signal after multi-scale filtering, the random error components in the registered image are quantized; (2) Structural Preservation – By calculating the low-frequency structural energy and local correlation in the difference plot, the continuity and fidelity of real biological structures are assessed.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A multi-scale differential evaluation method for real microscope image data includes the following steps: S1: Acquire microscope images and construct a difference image; S2: Enhance the intensity variation in the difference between adjacent slices using Laplace filtering; S3: Quantitatively describe noise characteristics from two aspects: entropy and signal-to-noise ratio (SNR); S4: Extract texture features at different spatial scales through multi-scale differential filtering (such as Gaussian difference pyramid) to reflect the continuity and fidelity of the structure; S5: Evaluation is performed using a combination of texture contrast, noise distribution, and information entropy characteristics in multi-scale images.

[0009] Furthermore, in S1, for two adjacent slices in the microscope image sequence... and First, calculate their difference image: , The difference image Used to reflect the change in grayscale intensity between adjacent slices; if the registration accuracy is insufficient, then The medium contains a large amount of unstructured high-frequency noise; if the registration is good, then... It presents a continuous texture distribution consistent with the structural changes of real organisms.

[0010] Furthermore, in S2, since nonlinear deformation often leads to pixel-level intensity abrupt changes, the Laplacian operator is used to perform high-pass enhancement on the difference map: , in, The Laplacian operator is used, where D represents the difference map, and x and y represent the horizontal and vertical gradient calculations. This operation can amplify local gray-level gradient changes, making random noise and fine structures distinguishable in the frequency domain.

[0011] Furthermore, in S3:

[0012] (1) Entropy calculation: The entropy of the Laplace filter result is used to describe the degree of randomness of the gray-level distribution:

[0013] Where H represents the calculated entropy, , Let i be the number of pixels with grayscale value i, and N be the total number of pixels in the image.

[0014] A higher entropy value indicates stronger uncertainty and noise in the image; (2) Signal-to-noise ratio calculation: To separate the structural signal from the noise, this invention uses a bilateral filter. Perform smooth decomposition:

[0015] Where Bilateralfilter represents a two-sided filter. Indicates the smoothed structural components. This represents the noise residual. The signal-to-noise ratio (SNR) is defined as:

[0016] in This represents the variance operator.

[0017] A higher SNR value indicates lower noise after registration and more complete preservation of structural information.

[0018] Furthermore, in S4: To evaluate the continuity and texture preservation of biological structures after registration, a Difference of Gaussian (DoG) pyramid was constructed based on the Laplacian difference image.

[0019] in For the i-th layer Gaussian kernel, This indicates that Gaussian smoothing is applied to the (x,y) position of the i-th Gaussian kernel. The variance of the Gaussian kernel is represented. This represents the Gaussian difference image of the i-th layer; The DoG pyramid suppresses noise and enhances edge features at different scales; Then, the local contrast of each layer is calculated to measure structural sharpness:

[0020] in For the first The gray-level co-occurrence matrix of the k-th and j-th pixels in a layer DoG image; higher This indicates that the texture of biological structures is clearer at this scale.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects: Compared with existing methods for evaluating the registration quality of microscope images, the multi-scale difference evaluation algorithm proposed in this invention has significant improvements in evaluation objectivity, stability, and adaptability.

[0022] First, this invention achieves objective and automated evaluation even without real-world ground images. Traditional methods rely on manual annotation or structural comparison, which are highly subjective and difficult to generalize. In contrast, this invention, through difference image analysis between adjacent slices and a multi-scale filtering strategy, can independently evaluate registration results without manual intervention, significantly improving evaluation efficiency and consistency.

[0023] Secondly, this invention comprehensively measures registration quality from two complementary perspectives: noise suppression and structural fidelity. By extracting the intensity changes of the difference map between adjacent slices through Laplacian filtering, this invention can sensitively capture the distribution of unstructured noise. Combined with bilateral filtering decomposition, it can calculate the signal-to-noise ratio (SNR) and entropy histogram, thereby quantitatively reflecting the noise level of the registered image. Simultaneously, this invention introduces a Gaussian difference (DoG) pyramid structure to suppress noise, enhance edges, and extract local contrast in multi-scale space, measuring the continuity and richness of biological texture features. In unregistered images, the Laplacian difference map shows significant random noise, while in registered images, cell boundaries and substructures are clearly preserved, and structural contrast is significantly improved at all scales.

[0024] Furthermore, the evaluation results of this invention have clear physical meaning and interpretability. The SNR index reflects the noise suppression capability after registration, the entropy value reflects residual randomness, and the multi-scale contrast quantifies structural fidelity. The combined use of these three metrics makes the registration quality evaluation more comprehensive, quantitative, and traceable, providing a scientific basis for subsequent automated reconstruction and algorithm optimization. This invention features high automation, strong robustness, and wide applicability, and can provide a reliable quality evaluation basis for three-dimensional reconstruction of microscope images, neuron tracking, and subsequent structural analysis. Attached Figure Description

[0025] Figure 1 A schematic diagram of the overall process of this invention.

[0026] Figure 2 This document presents the evaluation of various registration methods for this invention on three datasets. Figure a shows the contrast of each layer in the Gaussian pyramid corresponding to different scale factors, used to measure structural sharpness; higher contrast indicates sharper structure. Figure b shows the information entropy at different slice intervals. Figure c shows the signal-to-noise ratio at different slice intervals. Figure d shows the probability density distribution of pixel intensity. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description, in conjunction with specific embodiments and accompanying drawings, further illustrates the invention. Obviously, the described embodiments are only a portion, not all, of the embodiments disclosed in this invention. All other embodiments obtained by those skilled in the art based on the embodiments disclosed in this invention without inventive effort are within the scope of protection of this invention.

[0028] Example 1

[0029] The evaluation process of this invention mainly includes five stages: difference map calculation, Laplace enhancement, noise analysis, multi-scale structural analysis, and comprehensive evaluation. Figure 1 As shown.

[0030] (I) Construction of difference images For two adjacent slices in a microscope image sequence and First, calculate their difference image: , The difference image This is used to reflect the change in grayscale intensity between adjacent slices. If the registration accuracy is insufficient, then... The medium contains a large amount of unstructured high-frequency noise; if the registration is good, then... It presents a continuous texture distribution consistent with the structural changes of real organisms.

[0031] (ii) Laplace filter enhancement Since nonlinear deformation often leads to pixel-level intensity abrupt changes, this invention uses the Laplacian operator to perform high-pass enhancement on the difference map: , in This is the Laplacian operator. This operation amplifies local gray-level gradient changes, making random noise and fine structures distinguishable in the frequency domain.

[0032] (III) Noise Analysis and Quantification To quantitatively describe noise characteristics, this invention analyzes noise from two aspects: entropy and signal-to-noise ratio (SNR).

[0033] 1. Entropy Calculation: The entropy of the Laplace filter result is used to describe the degree of randomness in the gray-level distribution.

[0034] in , Let i be the number of pixels with grayscale value i, and N be the total number of pixels in the image. A larger entropy value indicates stronger uncertainty and noise in the image.

[0035] 2. Signal-to-noise ratio calculation: To separate the structural signal from the noise, this invention uses a bilateral filter for smooth decomposition:

[0036] in Indicates the smoothed structural components. This represents the noise residual. The signal-to-noise ratio (SNR) is defined as:

[0037] in This represents the variance operator. A higher SNR value indicates lower noise and more complete preservation of structural information after registration.

[0038] (iv) Multi-scale structural fidelity assessment To evaluate the continuity and texture preservation of biological structures after registration, this invention constructs a Difference of Gaussian (DoG) pyramid based on Laplacian difference images:

[0039] in For the i-th layer Gaussian kernel, This represents the Gaussian difference image of the i-th layer.

[0040] The DoG pyramid suppresses noise and enhances edge features at different scales. Subsequently, this invention calculates the local contrast of each layer to measure structural sharpness.

[0041] in For the first The gray-level co-occurrence matrix of a layer DoG image. Higher... This indicates that the texture of biological structures is clearer at this scale.

[0042] Example 2: To address the challenge of evaluating registration performance in real-world datasets lacking ground-based data, this invention proposes a multi-scale differential filtering evaluation strategy. This method achieves a balance between noise reduction and biological structure fidelity, enabling a comprehensive and robust evaluation of 3D registration results.

[0043] To verify its effectiveness, this embodiment conducted experiments on three ground-based simulated datasets on the OpenOrganelle platform. These datasets included high-resolution electron microscopy images of various mouse tissues, including the heart, kidney, and liver, allowing us to comprehensively evaluate the applicability and robustness of our method across different representative tissue types. Each dataset contained approximately 1000 slices, each with a resolution of 1184x1184 pixels. The benchmark test was conducted on a high-performance workstation equipped with an NVIDIA A100 GPU (40~GB of RAM), two AMD EPYC 7T83 CPUs (96 physical cores, 192 threads), and 512~GB of system memory. The measurement results correspond to a single-batch inference setting. Figure 2Evaluation results under different conditions are presented, including noisy data, SEAMLeSS, vEMRec, and ground reality, where SEAMLeSS and vEMRec are readily available registration algorithms. The results show that, in terms of noise reduction performance, ground reality has the lowest entropy and the highest signal-to-noise ratio (SNR), followed by vEMRec and SEAMLeSS, with unregistered noisy data lasting. Regarding the preservation of biological structural information, ground reality also exhibits the highest contrast across all scales, followed by vEMRec, while SEAMLeSS and noisy data perform similarly. This trend is consistent with quantitative results based on MI, NCC, and SSIM, where MI (mutual information) reflects the statistical dependence between images, NCC (normalized cross-correlation) measures the linear correlation of image brightness, and SSIM (structural similarity) assesses the consistency of images in terms of brightness, contrast, and structure. In summary, the experiments verified that the proposed multi-scale differential filtering evaluation strategy can effectively and reliably reflect the registration quality, and has high accuracy and robustness.

[0044] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale differential evaluation method for real microscope image data, characterized in that, Includes the following steps: S1: Acquire microscope images and construct a difference image; in S1, for two adjacent slices in the microscope image sequence... and First, calculate their difference image: , The difference image Used to reflect the change in grayscale intensity between adjacent slices; if the registration accuracy is insufficient, then The medium contains a large amount of unstructured high-frequency noise; if the registration is good, then... It presents a continuous texture distribution consistent with the structural changes of real organisms; S2: Enhance the intensity variation in the difference between adjacent slices using Laplace filtering; S3: Quantitatively describe noise characteristics from two aspects: entropy and signal-to-noise ratio; S4: Texture features are extracted at different spatial scales using multi-scale difference filtering to reflect the continuity and fidelity of the structure; to evaluate the continuity and texture preservation of the registered biological structure, a Gaussian difference pyramid (DoG) is constructed based on the Laplacian difference image. in, For the i-th layer Gaussian kernel, The variance of the Gaussian kernel is represented. This represents the Gaussian difference image of the i-th layer; the DoG pyramid suppresses noise and enhances edge features at different scales; Subsequently, the local contrast (Contrast) of each layer is calculated to measure structural sharpness: in For the first The gray-level co-occurrence matrix of the k-th and j-th pixels in the layer DoG image; S5: Evaluation is performed using a combination of texture contrast, noise distribution, and information entropy characteristics in multi-scale images.

2. The multi-scale difference evaluation method as described in claim 1, characterized in that, In step S2, since nonlinear deformation often leads to pixel-level intensity abrupt changes, the Laplacian operator is used to perform high-pass enhancement on the difference map: + , in, For the Laplace operator, Represents a difference plot. and This indicates the calculation of gradients in both the horizontal and vertical directions.

3. The multi-scale difference evaluation method as described in claim 1, characterized in that, In S3: (1) Entropy calculation: The entropy of the Laplace filter result is used to describe the degree of randomness of the gray-level distribution: Where H represents the calculated entropy, , Where i is the number of pixels with grayscale value i, and N is the total number of pixels in the image; the larger the entropy value, the stronger the uncertainty noise in the image; (2) Signal-to-noise ratio calculation: To separate the structural signal from the noise, a bilateral filter is used. Perform smooth decomposition: in Indicates a bilateral filter. Indicates the smoothed structural components. Represents the noise residual; the signal-to-noise ratio (SNR) is defined as: in This represents the variance operator.

Images