Organ-like three-dimensional image segmentation method and system and electronic equipment

Abstract

The invention relates to the field of biomedical image processing, and discloses an organ-like three-dimensional image segmentation method and system and electronic equipment, and the method comprises the steps: carrying out the slicing processing of the three-dimensional image data of an organ, converting the data into a two-dimensional image sequence, and carrying out the fine preprocessing of each two-dimensional slice, so as to eliminate the noise and enhance the image features; and then, adaptive time sequence enhancement processing is carried out, effective extraction and enhancement of time sequence features are carried out on the preprocessed two-dimensional image sequence, and the utilization rate of image information is improved. On this basis, a Kolmogorov-Arnold theory and a UNet architecture are combined to construct a high-performance image segmentation model, and deep training and testing of the model are carried out through an organ-like three-dimensional image data set. And finally, inputting the enhanced two-dimensional image sequence into the trained model to realize high-precision segmentation of the organoid. The method has the advantages of high segmentation accuracy, high processing efficiency and the like, and has relatively high practical value and application prospect.

Description

Technical Field

[0001] The present invention relates to the field of biomedical image processing, and in particular to a method, system and electronic equipment for segmenting organoid three-dimensional images. Background Art

[0002] Organoids have many key structural and functional characteristics of organs and tumor tissues, providing new technical means for disease modeling, personalized medicine, drug screening, and cell therapy. Organoids are located in matrix gel and self-assemble and grow in it, with multi-scale tissue morphology of 30-800μm (closely related to function and tumor heterogeneity). Three-dimensional imaging and accurate analysis of organoids are of great value for in-depth understanding of the growth and development of organoids and the effects of drugs.

[0003] Three-dimensional image segmentation of organoids is one of the key tasks in organoid morphological analysis, which aims to accurately extract the outline and internal structure of organoids from three-dimensional images. Due to the high heterogeneity of organoids in volume, morphology and distribution, and the interference of noise, signal discontinuity and other factors in the three-dimensional imaging process, traditional image segmentation methods often have difficulty in achieving high accuracy and robustness when processing organoids. Deep learning methods, especially image segmentation techniques based on convolutional neural networks (CNNs) and U-Net architectures, have gradually become the mainstream methods for organoid image segmentation. However, when imaging organoids using techniques such as optical coherence tomography (OCT), the signal intensity is insufficient, and the signal noise and intensity attenuation make the boundaries of organoids often unclear, and the segmentation results may not be closed, thus affecting subsequent analysis and diagnosis. Two-dimensional CNN and U-Net networks often ignore the temporal information between three-dimensional image frames, while three-dimensional CNN and U-Net networks have high requirements for computing resources and high pressure on data annotation. Therefore, how to improve segmentation accuracy while reducing computing resources, especially how to ensure the continuity and integrity of organoid boundaries, has become the key to improving the performance of three-dimensional image segmentation of organoids. Summary of the invention

[0004] The object of the present invention is to provide a method, system and electronic device for segmenting organoid three-dimensional images to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for segmenting a three-dimensional organoid image, the method comprising: The three-dimensional image data of the organoids were preprocessed and adaptively enhanced in time series. The enhanced image data were input into the image segmentation model based on the integration of Kolmogorov–Arnold theory and UNet architecture to output the segmentation results of the organoids.

[0006] Preferably, the adaptive timing enhancement process includes: The signal intensity of each frame in the input organoid two-dimensional image sequence is evaluated, the weight of each frame is dynamically adjusted, the similarity between the current frame and its adjacent frames is calculated, and adjacent frames with similarity to the current frame higher than a threshold are selected for weighted average fusion to generate an enhanced current frame image. The same processing is performed on all frames until the extraction and enhancement of time series features are completed.

[0007] Preferably, the signal strength evaluation method is: converting image data into amplitude data, calculating the amplitude logarithm of each pixel, converting the signal strength into decibel units, and assigning frame weights according to the signal strength.

[0008] Preferably, the method for calculating the similarity is: using the Bhattacharyya coefficient of the color histogram, the Euclidean distance of the texture feature based on the gray level co-occurrence matrix and the depth difference metric, and combining these similarities to obtain the final similarity.

[0009] Preferably, the image segmentation model applies a Kolmogorov–Arnold activation function to perform nonlinear transformation on the vectorized two-dimensional image features, and improves the diversity of features through recursive calculation and multi-level mapping.

[0010] Preferably, the image segmentation model inputs the features processed by the Kolmogorov–Arnold layer into a deep convolutional layer, further extracts features through the deep convolutional layer, and fuses features from different stages through a global attention mechanism and jump connections.

[0011] Preferably, pixel-level cross entropy loss function and Dice loss are used to optimize model parameters during the model training stage.

[0012] Preferably, before preprocessing and adaptive timing enhancement of the three-dimensional image data of the organoid, the organoid image is subjected to multi-scale slice processing by adjusting the slice thickness and number and converted into a two-dimensional image sequence to enhance the image details, significantly improve the accuracy and consistency of organoid segmentation, and effectively solve the problem of detection discontinuity caused by insufficient signal intensity. Preferably, the preprocessing includes normalizing, contrast enhancing and resizing the two-dimensional image sequence.

[0013] Wherein, the normalization processing includes converting the pixel values ​​of the image into a specified value range by using a normalization formula; The contrast enhancement includes using a linear stretching method to adjust the pixel value range of the image, a histogram equalization technique to adjust the pixel distribution of the image, an adaptive histogram equalization technique to adjust the local features of the image, and an edge detection algorithm Canny operator to enhance the edge information of the organoid image; The resizing includes interpolating the image and using cropping techniques to maintain key areas of the organoid in the image.

[0014] The present invention also discloses a three-dimensional organoid image segmentation system, including a system for executing the above-mentioned three-dimensional organoid image segmentation method.

[0015] The present invention also discloses an electronic device, comprising: a processor; and a memory, which stores executable program instructions. When the program instructions are executed by the processor, the device implements the above-mentioned organoid three-dimensional image segmentation method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Through adaptive temporal enhancement processing, the segmentation results are made more delicate and accurate in the three-dimensional structure processing of organoid images. The adaptive temporal enhancement learning mechanism is introduced to optimize different image features to ensure that when processing images with different signal intensities, the parameters can be automatically adjusted to enhance the segmentation effect, thereby coping with the discontinuity problem in the organoid image and improving the closure and continuity of the segmentation results. The use of the adaptive temporal enhancement learning mechanism enables better adaptability and generalization in the processing of diversified image data, reducing the complexity of artificial parameter adjustment. The present invention has strong nonlinear modeling capabilities. The structure and morphology of organoids are usually highly irregular. Traditional convolutional neural network (CNN) methods may not be able to fully capture their details when processing such complex data. The present invention can better model the complex spatiotemporal changes in organoid data by introducing the Kolmogorov–Arnold activation function, through recursive calculation and the introduction of nonlinear activation functions, so that the network can more accurately model the complexity of organoids at multiple levels, especially when processing subtle features with dynamic changes and multidimensional correlations in organoids, and improve the ability to segment the internal details and boundaries of organoids.

[0017] The present invention designs a preprocessing step based on edge detection to further optimize the segmentation results, ensure the closure and continuity of the organoid boundaries, reduce errors caused by insufficient signals, improve the clinical application value of the segmentation results, and better support the implementation of personalized medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A processing flow chart of a method for segmenting a three-dimensional organoid image according to the present invention; Figure 2A schematic diagram of the design for adaptive timing enhancement processing; Figure 3 It is a structural diagram of the image segmentation model described in the present invention; Figure 4 The processing flow chart of image contrast enhancement; Figure 5 Processing flow chart for image resizing; Figure 6 A schematic diagram of a three-dimensional organoid image segmentation system according to the present invention; Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1-5 The present invention provides a technical solution: a method for segmenting a three-dimensional organoid image, the method comprising: Acquire three-dimensional image data of organoids, which usually exist in the form of voxels, representing points in three-dimensional space. Then, use image processing technology to slice the three-dimensional volume data along a certain dimension to obtain a series of two-dimensional image sequences. Each two-dimensional slice image represents the morphological structure of the organoid at that specific depth. The obtained two-dimensional slice images are preprocessed to improve the image quality and provide clearer and more accurate input data for subsequent processing. Adaptive time series enhancement processing is performed on the two-dimensional slice images to extract and enhance the time series features in the two-dimensional image sequence. Specifically, the preprocessed two-dimensional image sequence is analyzed frame by frame to capture the time dependency and dynamic changes between images, so as to achieve effective feature extraction and enhancement.

[0021] The enhanced image data was input into the image segmentation model based on Kolmogorov–Arnold theory and UNet network architecture. Kolmogorov–Arnold theory provides a theoretical basis for the model, ensuring that the model can approximate complex nonlinear functional relationships; and the UNet architecture was selected as the basic framework of the model due to its excellent image segmentation performance. Subsequently, the model was trained using a labeled organoid 3D image dataset, and the model parameters were adjusted through the back propagation algorithm and performance optimization algorithm, so that the model achieved a high segmentation accuracy on the training set.

[0022] The two-dimensional image sequence after adaptive temporal enhancement is input into the trained image segmentation model. The model segments the input image sequence frame by frame and finally obtains the segmentation result of the organoid in each two-dimensional slice image.

[0023] The present invention will be further described below in conjunction with Examples 1 to 3: Embodiment 1: Additionally or optionally, this embodiment is used to describe a method for implementing conversion of three-dimensional volume data into a two-dimensional image sequence, and the specific implementation includes: Load medical 3D image data and use the nibabel library to load medical image data in .nii.gz format. The data is stored as a 3D array, usually in shape (D, H, W), where D is the depth, H is the height, and W is the width.

[0024]

[0025] Determine the slice direction and select the slice axis according to the view direction specified by the user. For the x direction, select axis 0; for the z direction, select axis 1; for the y direction, select axis 2. Generate slices and extract each slice by traversing the selected axis.

[0026] After acquiring the two-dimensional image sequence, further image preprocessing is required, and the preprocessing methods include normalization, contrast enhancement and size adjustment.

[0027] In order to ensure the quality of image data during visualization, each slice is normalized to convert the pixel values ​​of the slice image to a specified numerical range. The normalization formula is:

[0028] in, Represents the normalized image data, slice_data represents the original slice data, min(slice_data) represents the minimum value in the slice data, max(slice_data) represents the maximum value in the slice data, and the value 255 represents the upper limit of the normalized data range, which is used to convert the data to uint8 type.

[0029] Since organoid images usually have low contrast in OCT, the boundaries of organoids may not be clear enough. In order to improve the recognition and segmentation accuracy of images, the present invention adopts a contrast enhancement method to highlight the important features of organoids and improve the recognizability of image details.

[0030] The specific operation is as follows: First, the pixel value range of the image is adjusted by the linear stretching method, and the low-contrast part of the image is stretched to a higher contrast range. The specific method is to map the pixel value to the preset high and low contrast ranges to enhance the brightness contrast of the image. Next, the image is processed by the histogram equalization technology, and the brightness range of the image is made more uniform by adjusting the pixel distribution of the image. This method is particularly suitable for dark or bright areas in the image, and can effectively improve the overall contrast of the image and highlight the boundaries of the organoids. Then, the adaptive histogram equalization technology is used. This method avoids noise amplification caused by excessive contrast enhancement through block area processing, and adaptively enhances the contrast of the image, especially in the local area of ​​the image (such as the edge of the organoid) to enhance the detail display and highlight the target area. Finally, the edge information of the organoid image is enhanced by the edge detection algorithm Canny operator. This process further improves the accuracy of its segmentation by strengthening the edges of the organoids in the image, and provides clearer feature information for subsequent models.

[0031] Since organoid data usually have different resolutions and sizes, and have irregular morphological changes, unified resizing is required to ensure that the model can accurately process organoid images of different scales. Organoid images are resized based on their three-dimensional structural characteristics.

[0032] The specific operations are as follows: First, according to the model input requirements, a standard size is selected. The target size should be reasonably set according to the typical distribution and characteristics of organoid images to ensure that image details are not lost. When resizing, the aspect ratio of the image is first calculated and the image is interpolated. In order to avoid morphological distortion of the organoid, methods such as bilinear interpolation or cubic interpolation are used to resize the image in a smooth transition manner to ensure that important details are not compressed or stretched. In some cases, organoid images may contain unnecessary background areas that may interfere with the segmentation process. To this end, adaptive cropping techniques are used to crop according to the boundaries and morphological features of the organoids, remove redundant background, and maintain the key areas of the organoids in the image. Finally, during the adjustment process, special attention is paid to the structural details of the organoids to ensure that their key features can still be clearly presented after resizing. For smaller organoid images, local enhancement strategies may be used to maintain the clarity of important details.

[0033] Generate and save the image. Use the PIL library to convert the normalized slices to image format and perform necessary rotations based on the specified view direction. Finally, save the image to the specified output directory.

[0034] Embodiment 2: Additionally or optionally, due to certain discontinuities in the OCT image imaging process, some areas of a single image may have weak or unclear imaging signals after slicing. This discontinuity poses a challenge to the precise segmentation of organoids. In order to solve this problem, the present invention adopts an adaptive time series enhancement processing method, which makes up for the missing or weakened signal in a single image by fully exploring the characteristics of the image in the time series. By enhancing the time series information, the segmentation effect of organoids can be effectively improved, ensuring the accuracy and stability of image segmentation.

[0035] like Figure 2 As shown in the figure, the core function of adaptive temporal enhancement is to enhance the perception of dynamic changes by learning the information of previous and next frames. Logically, by analyzing the temporal dependency of previous and next frames, linear temporal features can be effectively extracted. Linear features help capture smooth changes. Therefore, adaptive temporal enhancement can not only improve the recognition ability of static structures, but also enhance the response ability to temporal changes, thereby adapting to the complex temporal nature of organoid data. Adaptive temporal enhancement specifically includes: The signal strength of the input OCT image is evaluated, mainly by adjusting the weight according to the signal strength of each frame of the image, and frames with high signal strength (such as frames with clear edges or obvious target areas) are given higher weights. The signal strength of each frame of the image is calculated, and frames with higher signal strength represent frames that contain more effective information. Specifically, the image data is converted into amplitude data, and then the signal strength is calculated. The logarithm of the amplitude of each pixel is taken (usually the logarithm with a base of 10) to convert the signal strength into decibel (dB) units. The weight of each frame is dynamically adjusted according to the signal strength. Frames with higher signal strength can be given higher weights, so that these frames will have a greater impact in the subsequent adaptive timing enhancement processing. Then the image is reconstructed to reconstruct the calculated signal strength data into a two-dimensional image, that is, the signal strength data is mapped to the grayscale value of the image.

[0036] The purpose of calculating the similarity between the current index frame and the adjacent frames is to calculate the similarity between the adjacent frames and the current frame. Frames with high similarity are given higher weights. By evaluating the similarity between adjacent frames and the current frame, helpful frames can be selected in a targeted manner. Specifically, inter-frame similarity evaluation is first performed, and the similarity between adjacent frames and the current frame is calculated using image features (color histogram, texture features, or depth features).

[0037] Color histogram is a common method to describe the color distribution of an image. and , the image is a grayscale image, and the similarity of the grayscale histogram can be measured using the Bhattacharyya coefficient:

[0038] in, and They are images and Grayscale histogram of , N is the total number of gray levels.

[0039] Texture features are usually described using a gray-level co-occurrence matrix (GLCM). Statistics such as contrast, correlation, and energy calculated based on GLCM can be used as indicators of similarity. The similarity metric used is the Euclidean distance:

[0040] in, and They are images and Vector representation of texture features, ||h|| represents the L2 norm of the vector.

[0041] Depth images can be used to measure similarity by calculating the matching of depth point clouds. Use the depth difference metric SAD (Sum of Absolute Differences):

[0042] here, and is an image and The depth value of the corresponding pixel, N is the total number of pixels in the image.

[0043] Three similarities are calculated by color histogram, texture and depth features. Next, these similarities are combined to get the final similarity:

[0044] in, are weight coefficients used to balance the contribution of different features to the final similarity. These coefficients can be adjusted according to the required feature importance. Finally, the comprehensive similarity Converted into weights and applied to subsequent image processing or fusion.

[0045] Select frames based on similarity. Based on the calculated similarity, select adjacent frames with higher similarity to the current frame. These frames are given higher weights, thereby enhancing the timing effect. Specifically, convert similarity into weights. In order to strengthen the influence of certain highly similar frames, an exponential weighting method is used:

[0046] Among them, S is the similarity between a pair of frames, N is the similarity set between all frames, and w is the normalized weight.

[0047] Perform adaptive frame selection. Automatically select the most relevant adjacent frames based on the features of the current frame, rather than a fixed number of frames. In traditional methods, a fixed number of adjacent frames are usually selected for processing, while adaptive frame selection dynamically selects the most relevant frames based on the features of each frame. This ensures that only the frames most relevant to the current frame are selected, while avoiding the redundant information that comes with a fixed number of frame selections. Specifically, features are first extracted for each frame (such as areas containing organoids and matrix gel), and the importance of each frame is evaluated based on these features. Next, dynamic frame selection is performed to adaptively select highly correlated adjacent frames for temporal enhancement based on the features of the current frame, rather than selecting a fixed number of frames.

[0048] The purpose of using the dynamic number selection method is to be more flexible in selecting the number of frames, and to dynamically select the number of frames based on the distribution of similarities. A dynamic threshold or number k is selected based on the similarity distribution between the current frame and the adjacent frames. Specifically, after calculating the similarity between all adjacent frames and the current frame, the average μ and standard deviation σ of these similarities are calculated. Next, the selection criterion is set to select frames with a similarity greater than μ+λσ, where λ is a regulation factor used to control the number of selected frames:

[0049] Among them, μ and σ are the mean and standard deviation of similarity, respectively, and λ is the adjustment factor. This method can dynamically adjust the number of selected frames according to the similarity distribution.

[0050] After the signal strength evaluation, similarity calculation and adaptive frame selection of the current indexed image, the frame number index set to be selected can be obtained. and the weight set assigned to the selected frame Assume that the current image is X and the current index is current. To calculate the enhanced image, the following formula is used to combine the current image and the images of the previous and next frames: This formula integrates the information of the current image and adjacent frame images into an enhanced image by weighted averaging.

[0051] Embodiment 3: Additionally or optionally, the present embodiment is an implementation method for segmenting organoid images using an image segmentation model. In previous organoid segmentation tasks, the inherent ambiguity of the multi-layer perceptron (MLP) structure limits the interpretability of the model, making it difficult to understand and explain its decision-making process. In addition, MLP typically relies on a linear weight matrix, which may limit the performance and structural design of the model. Kolmogorov–Arnold gets rid of the reliance on linear weight matrices in traditional MLP by innovatively using learnable activation functions at the edge of the network and using these parameterized activation functions as weights. This design not only reduces the size of the model, but also improves the interpretability of the model while maintaining or exceeding the original performance.

[0052] More importantly, the essence of the Kolmogorov–Arnold method when processing time series data is to model nonlinear dynamic systems. Unlike traditional methods, the Kolmogorov–Arnold method emphasizes the inherent complexity of time series data and reveals the underlying laws behind the data through embedding mapping in high-dimensional space. This innovation is the ability to capture tiny time series changes in high-dimensional space, which is particularly suitable for processing complex nonlinear time series relationships in organoid data.

[0053] Through this modeling method, the Kolmogorov–Arnold method significantly improves the model's understanding and prediction capabilities of spatiotemporal features, enabling it to more effectively adapt to the complex data features in organoid segmentation tasks, especially in scenes with dynamic changes and strong multi-dimensional correlations. The specific methods for building image segmentation models include: like Figure 3 As shown in Figure 1, first, input data processing and multi-scale feature extraction. Image I is used as network input. , H and W represent the length and width of the image respectively, and C represents the number of image channels. Multi-scale feature extraction is performed through convolution kernels of different sizes to obtain multi-scale image features, thereby providing more dimensional information for subsequent processing stages.

[0054] The image passes through a series of convolutional layers (Conv), batch normalization layers (BN), ReLU activation functions, and maximum pooling layers to generate the first Feature map of the layer , there are L such configurations, and the final output is derived as At this stage, multi-scale convolutional features are fused to obtain a more comprehensive image representation. The convolution kernel size and stride of each layer are determined by a dynamic optimization strategy to automatically adjust the network structure according to the size of the image.

[0055] The output features of the convolution stage Reshape into a series of 2D tiles, each of size , the total number of patches is These patches are vectorized to obtain z, and a high-dimensional embedding space is constructed to capture the local and global temporal changes in the image, and it is mapped to a potential D-dimensional embedding space through a trainable linear projection matrix. In this process, an adaptive mechanism is used to adjust the activation function, so that the network can dynamically select the appropriate nonlinear transformation according to the characteristics of the input data.

[0056] The Kolmogorov–Arnold method is applied to the reshaped patches, and the vectorized elements z are input at each layer, and the input is nonlinearly mapped through a learnable activation function. The activation function of each layer can dynamically adjust its behavior according to different inputs through an adaptive mechanism. This structure allows each input to be mapped to the output through a specific set of activation functions without relying on the traditional linear weight matrix. Recursive calculations are introduced between layers so that each layer not only depends on the output of the previous layer, but also takes into account the global influence of all activation functions before the layer. This recursive calculation method makes the output of each layer more diversified, no longer relying solely on the linear weighted sum, and gives full play to the advantages of nonlinear activation functions.

[0057] The features processed by the Kolmogorov–Arnold layer are passed to the deep convolution layer (DwConv) and further processed by batch normalization (BN) and ReLU activation function. At this time, the newly added cross-layer connection not only includes the traditional residual connection, but also introduces a global attention mechanism to enhance the focus on key areas. Through this mechanism, the model can adaptively weight the key information in the image to improve segmentation accuracy.

[0058] Features from different stages are spliced ​​using skip connections and passed to the upsampling block. While skipping connections, the present invention introduces an adjustable upsampling stride and a nonlinear interpolation algorithm, so that the upsampling process not only improves the spatial resolution, but also effectively retains the details of low-level features. In order to further extract higher-level features, the image is fused layer by layer through a series of convolutional layers. The convolution operation of each layer not only reduces the spatial dimension, but also enhances the semantic representation of the image through a nonlinear activation function. In this process, by introducing a multi-scale feature fusion strategy, features of different scales can be effectively fused, helping the model retain key information when processing complex structures.

[0059] The last layer of the convolutional layer is a 1×1 convolution, which is used to map each pixel to the classification space and output the predicted probability of each pixel in each category. This output is activated by the Softmax function to generate the probability distribution of each pixel corresponding to the category, ensuring that the predicted value of each pixel can be expressed as the probability of belonging to each category.

[0060] Finally, morphological operations are applied to post-process the output to remove noise and fix small defects in the segmentation results. The final organoid segmentation map is generated by thresholding the classification results. The output passes through a Softmax activation function so that the predicted value of each pixel represents the probability of belonging to each category. For each pixel i, its predicted probability of belonging to category c is .

[0061] During the training phase, pixel-level cross entropy loss and Dice loss are used to optimize model parameters. In the organoid semantic segmentation task, the final segmentation map can be output from the last layer. obtained in is the number of semantic categories and Y represents the ground truth segmentation. Therefore, the segmentation loss is the pixel-wise cross entropy loss and can be expressed as:

[0062] Where N is the number of training samples, C is the number of categories, is the indicator function of whether the i-th pixel belongs to the c-th class, is the predicted probability that the i-th pixel belongs to the c-th class.

[0063] Dice loss is used to measure the overlap between the predicted results and the true labels, which is particularly effective when dealing with class imbalance problems. The Dice loss function is defined as follows:

[0064] The total loss function combines the cross entropy loss and the Dice loss in a weighted manner to balance the classification accuracy and segmentation accuracy of the model:

[0065] in, and is the weight coefficient of cross entropy loss and Dice loss. By adjusting these two weight parameters, the model can control the emphasis on different losses, thereby achieving a better segmentation effect.

[0066] In the actual inference stage of the model, the model receives new organoid image data, and after processing through the above steps, it outputs an accurate segmentation mask, thereby achieving efficient and accurate segmentation of organoids.

[0067] Additionally or alternatively, such as Figure 6 As shown, the present invention also provides an organoid three-dimensional image segmentation system, including a system for executing the organoid three-dimensional image segmentation method described in the above embodiment, and also including: an adaptive timing enhancement module 101, used to perform adaptive timing enhancement processing on the three-dimensional image data of the organoid; The image segmentation module 102 constructs an image segmentation model based on the integration of Kolmogorov–Arnold theory and UNet architecture, and is used to receive enhanced image data and output organoid segmentation results.

[0068] Additionally or alternatively, such as Figure 7 As shown, the present invention further provides an electronic device 200, including: Processor 201; and The memory 202 stores executable program instructions. When the program instructions are executed by the processor, the device implements the organoid three-dimensional image segmentation method described in the above embodiment.

[0069] It is understood that in order to clearly illustrate the solution of the present disclosure and avoid confusion with the prior art, Figure 7 The electronic device 200 of the present disclosure only shows the components related to the embodiment of the present disclosure, and omits the components that may be necessary for implementing the embodiment of the present disclosure but belong to the scope of the prior art. Therefore, based on the content disclosed in the present disclosure, a person skilled in the art can clearly understand that the electronic device 200 of the present disclosure can also include components related to the embodiment of the present disclosure. Figure 7 The constituent elements shown in are different from the common constituent elements.

[0070] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0071] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for segmenting a three-dimensional organoid image, characterized in that: The method comprises: The three-dimensional image data of the organoids were preprocessed and adaptively enhanced in time series. The enhanced image data were input into the image segmentation model based on the integration of Kolmogorov–Arnold theory and UNet architecture to output the segmentation results of the organoids.

2. The method for segmenting an organoid three-dimensional image according to claim 1, characterized in that: The adaptive timing enhancement process comprises: The signal intensity of each frame in the input organoid two-dimensional image sequence is evaluated, the weight of each frame is dynamically adjusted, the similarity between the current frame and its adjacent frames is calculated, and adjacent frames with similarity to the current frame higher than a threshold are selected for weighted average fusion to generate an enhanced current frame image. The same processing is performed on all frames until the extraction and enhancement of time series features are completed.

3. The method for segmenting an organoid three-dimensional image according to claim 2, characterized in that: The signal strength evaluation method is: converting image data into amplitude data, calculating the amplitude logarithm of each pixel, converting the signal strength into decibel units, and assigning frame weights according to the signal strength.

4. The method for segmenting an organoid three-dimensional image according to claim 2, characterized in that: The method for calculating the similarity is: using the Bhattacharyya coefficient of the color histogram, the texture feature based on the gray level co-occurrence matrix, the Euclidean distance and the depth difference metric, and combining these similarities to obtain the final similarity.

5. A method for segmenting an organoid three-dimensional image according to any one of claims 1 to 4, characterized in that: Before preprocessing and adaptive temporal enhancement of the three-dimensional image data of the organoid, the organoid images are converted into two-dimensional image sequences; The preprocessing includes normalizing, contrast enhancing and resizing the two-dimensional image sequence.

6. The method for segmenting an organoid three-dimensional image according to claim 5, characterized in that: The normalization process includes converting the pixel values ​​of the image into a specified value range by using a normalization formula; The contrast enhancement includes using a linear stretching method to adjust the pixel value range of the image, a histogram equalization technique to adjust the pixel distribution of the image, an adaptive histogram equalization technique to adjust the local features of the image, and an edge detection algorithm Canny operator to enhance the edge information of the organoid image; The resizing includes interpolating the image and using cropping techniques to maintain key areas of the organoid in the image.

7. A method for segmenting an organoid three-dimensional image according to any one of claims 1 to 4, characterized in that: The image segmentation model applies the Kolmogorov–Arnold activation function to the vectorized two-dimensional image features for nonlinear transformation, and improves the diversity of features through recursive calculation and multi-level mapping; The image segmentation model inputs the features processed by the Kolmogorov–Arnold layer into the deep convolution layer, further extracts the features through the deep convolution layer, and fuses the features of different stages through the global attention mechanism and jump connection.

8. A method for segmenting an organoid three-dimensional image according to any one of claims 1 to 4, characterized in that: Also includes: During the model training phase, pixel-level cross entropy loss function and Dice loss are used to optimize model parameters.

9. A three-dimensional organoid image segmentation system, characterized in that: The method comprises a method for executing the organoid three-dimensional image segmentation method according to any one of claims 1 to 8.

10. An electronic device, characterized in that: include: processor; and a memory storing executable program instructions, which, when the program instructions are executed by the processor, enable the device to implement the method according to any one of claims 1 to 8.

Images