Micro-hyperspectral gallbladder cancer image dimension reduction method

Abstract

The application discloses a kind of micro-hyperspectral gallbladder cancer image dimension reduction method, can reduce the feature redundancy of hyperspectral gallbladder cancer image, to efficiently extract the feature of hyperspectral gallbladder cancer image, significantly improve the accuracy of cancerous site classification, and dimension reduction effect is compared with traditional PCA and LDA etc., has obvious advantage on hyperspectral gallbladder cancer image classification.Compared with the prior art, the application has the advantages that traditional dimension reduction method (such as PCA and LDA etc.) is linear dimension reduction method, which can cause different sample categories to be inseparable.The hyperspectral gallbladder cancer image dimension reduction method based on superpixel manifold learning can effectively express the characteristics of high-dimensional data by introducing AWF filtering module, popular learning PaCMAP module, superpixel module and kernel principal component analysis module, while also compressing the data volume, which is more conducive to the rapid extraction of information.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral gallbladder cancer image classification, and specifically to a method for dimensionality reduction of microscopic hyperspectral gallbladder cancer images. Background Technology

[0002] Gallbladder carcinoma (GBC) is a common malignant tumor of the biliary system. It is highly malignant, difficult to diagnose early, and 70%–80% of patients are diagnosed at an advanced stage. Comprehensive treatments such as radiotherapy and chemotherapy are often ineffective, resulting in a very poor prognosis with a 5-year survival rate of only 5%–15%. Therefore, if radical resection can be performed in the early (T1) tumor stage, the patient's prognosis can be significantly improved, with a 5-year survival rate reaching 75%. Thus, early identification of early-stage gallbladder cancer and precancerous lesions will effectively improve clinical survival rates. Strengthening basic research on gallbladder cancer to improve its clinical diagnosis and treatment is key to improving the prognosis of gallbladder cancer patients.

[0003] Hyperspectral imaging technology uses hyperspectral sensors, or imaging spectrometers, mounted on different space platforms to simultaneously image a target region across dozens to hundreds of continuous and subdivided spectral bands in the ultraviolet, visible, near-infrared, and mid-infrared regions of the electromagnetic spectrum. It acquires both spatial and spectral information, achieving a combination of spectral and image data. Therefore, the impact and development potential of hyperspectral imaging technology are unparalleled by previous stages of technological development, attracting significant interest in the medical field.

[0004] Hyperspectral gallbladder cancer images contain rich spectral information; hundreds of continuous spectral bands allow us to distinguish cancerous and non-cancer regions with greater detail. However, due to the limited number of training samples, the high dimensionality of hyperspectral gallbladder cancer images increases computational complexity and may even reduce classification accuracy. Furthermore, the raw spectra alone are insufficient to accurately represent or reveal the intrinsic geometric structure of gallbladder cancer images. Therefore, dimensionality reduction for feature extraction is a crucial step in the classification of high-dimensional hyperspectral image data.

[0005] Linear feature extraction methods based on spectral features include Principal Component Analysis (PCA) and Independent Principal Component Analysis (IPA), while nonlinear manifold learning algorithms include Isometric Feature Mapping (ISOMAP), Local Linear Embedding (LLE), and Laplacian Feature Mapping (LE). Typical linear manifold learning algorithms include Local Preserving Projection (LPP) and Nearest Neighbor Preserving Embedding (NPE), which are linear extensions of LE and LLE algorithms, respectively. However, these methods only consider local or global features during dimensionality reduction, failing to consider both simultaneously. How to better consider both local and global features in image dimensionality reduction is a key challenge for hyperspectral gallbladder cancer image classification. Summary of the Invention

[0006] Purpose of the invention: To overcome the problem in existing hyperspectral gallbladder cancer image classification algorithms that lead to the indistinguishability of different sample categories, this invention provides a dimensionality reduction method for microscopic hyperspectral gallbladder cancer images. By introducing an AWF filtering module, a manifold learning PaCMAP module, a superpixel module, and a kernel principal component analysis module, this method can effectively represent the features of high-dimensional data using low-dimensional data by acquiring rich spatial-spectral features, while also compressing the data volume and facilitating faster information extraction.

[0007] Technical Solution: To achieve the above objectives, this invention provides a method for dimensionality reduction of microscopic hyperspectral gallbladder cancer images, comprising the following steps:

[0008] S1: PaCMAP technology is used to reduce the spectral dimension of hyperspectral images. The spectral dimension of the original dataset in the experiment is compressed to 3 dimensions using the PaCMAP method. While retaining effective spectral information, the amount of data is greatly reduced, thereby reducing the training time. Then, the AWF filter is used to filter the dimensionality-reduced data to smooth the pixels.

[0009] S2: Based on the set number of superpixel segments, a simple iterative linear clustering method is used to segment the filtered data to obtain several superpixel blocks. Kernel principal component analysis is then used to reduce the dimensionality of each superpixel block, yielding low-dimensional subspace features.

[0010] S3: The training samples in these low-dimensional image patches will be fed into the network model of the support vector machine model for model training;

[0011] S4: After the support vector machine is trained on samples, the test samples are input into the network to classify the test samples, and the final classification label is generated by the maximum voter.

[0012] 2. According to the aforementioned method for dimensionality reduction of microscopic hyperspectral gallbladder cancer images, the process of extracting features using PaCMAP in step S1 is characterized as follows:

[0013] A1: Find a certain number of near-range pixel pairs, mid-range pixel pairs, and long-range pixel pairs in the image respectively;

[0014] A2: Use PCA to initialize the embedding values ​​corresponding to the pixels;

[0015] A3: The loss function value is reduced to the minimum in three stages, and the final category value corresponding to the pixel is obtained.

[0016] 3. According to the aforementioned method for dimensionality reduction of microscopic hyperspectral gallbladder cancer images, the simple iterative linear clustering process in step S2 is characterized by:

[0017] B1: Initialize seed points (cluster centers): Distribute seed points evenly within the image according to the set number of superpixels. Assuming the image has a total of N pixels, pre-segmented into K superpixels of the same size, then the size of each superpixel is N / K, and the distance (step size) between adjacent seed points is approximately S = sqrt(N / K);

[0018] B2: Reselect a seed point within the n*n neighborhood of the seed point (usually n=3). The specific method is: calculate the gradient values ​​of all pixels in the neighborhood, and move the seed point to the location with the smallest gradient in the neighborhood;

[0019] B3: Assign a class label (i.e., which cluster center each pixel belongs to) within the neighborhood of each seed point. Unlike standard k-means, which searches the entire image, SLIC limits the search range to 2S*2S, which can accelerate the convergence of the algorithm.

[0020] B4: Distance metric. Includes color distance and spatial distance. For each found pixel, calculate its distance to the seed point.

[0021] B5: Iterative optimization. Theoretically, the above steps are iterated until the error converges (which can be understood as the cluster center of each pixel no longer changing);

[0022] B6: Enhance Connectivity. The above iterative optimization may result in the following drawbacks: multi-connectivity issues, excessively small superpixel sizes, and single superpixels being fragmented into multiple discontinuous superpixels. These issues can be addressed by enhancing connectivity. The main idea is: create a new label table with all elements set to -1. Following a "Z" shaped pattern (from left to right, from top to bottom), discontinuous superpixels and excessively small superpixels are reassigned to neighboring superpixels. Pixels that have been traversed are assigned their corresponding labels until all points have been traversed.

[0023] Beneficial effects: By introducing the AWF filtering module, the popular learning PaCMAP module, the superpixel module, and the kernel principal component analysis module, we can effectively express the features of high-dimensional data by acquiring rich spatial-spectral features and using low-dimensional data. At the same time, the amount of data is compressed, which is more conducive to the rapid extraction of information and ensures good and stable classification results. Attached Figure Description

[0024] Figure 1 This is a network model diagram for an embodiment of the present invention; Detailed Implementation

[0025] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0026] This invention provides a method for dimensionality reduction of microscopic hyperspectral gallbladder cancer images. First, an overall network is constructed, including an AWF filtering module, a manifold learning PaCMAP module, seven simple linear iterative clustering modules, and seven kernel principal component analysis modules, as detailed below. Figure 1 As shown, the AWF filtering module removes sharp noise from the original image, making the image features more obvious. Subsequently, the popular learning PaCMAP module extracts image features by combining pixel information from near, mid, and far images. Then, a simple linear iterative clustering module uses a small number of superpixels to replace a large number of pixels to express image features, reducing the complexity of image post-processing. Finally, a support vector machine model is used for classification, and the final classification result is obtained by the maximum voter.

[0027] Based on the above method, this embodiment applies the method to hyperspectral gallbladder cancer image classification, and the specific process is as follows:

[0028] Step 1: The spectral dimension of the hyperspectral image is reduced using the PaCMAP technique. The spectral dimension of the original dataset in the experiment is compressed to 3 dimensions using the PaCMAP method. This greatly reduces the amount of data while retaining effective spectral information, thereby reducing training time. Then, the AWF filter is used to filter the dimensionality-reduced data to smooth the pixels.

[0029] Step 2: Based on the set number of superpixel segments, a simple iterative linear clustering method is used to segment the filtered data to obtain several superpixel blocks. Kernel principal component analysis is then used to reduce the dimensionality of each superpixel block, yielding low-dimensional features in the low-dimensional subspace.

[0030] Step 3: The training samples from these low-dimensional image patches will be fed into the network model of the support vector machine model for model training;

[0031] Step 4: After the support vector machine has been trained on samples, the test samples are input into the network to classify the test samples, and the final classification label is generated by the maximum voter.

[0032] The popular PaCMAP feature extraction network learning process includes the following sub-steps:

[0033] Step 1.1: Find a certain number of near-range pixel pairs, mid-range pixel pairs, and long-range pixel pairs in the image;

[0034] Step 1.2: Initialize the embedding values ​​corresponding to the pixels using PCA;

[0035] Step 1.3: Reduce the value of the loss function to the minimum in three stages and obtain the final category value corresponding to the pixel.

[0036] The learning process of a simple iterative linear clustering network includes the following sub-steps:

[0037] Step 2.1: Initialize seed points (cluster centers): Distribute seed points evenly within the image according to the set number of superpixels. Assuming the image has a total of N pixels, pre-segmented into K superpixels of the same size, then the size of each superpixel is N / K, and the distance (step size) between adjacent seed points is approximately S = sqrt(N / K);

[0038] Step 2.2: Reselect a seed point within the n*n neighborhood of the seed point (generally n=3). The specific method is: calculate the gradient values ​​of all pixels within this neighborhood, and move the seed point to the location with the smallest gradient within that neighborhood;

[0039] Step 2.3: Assign a class label (i.e., which cluster center each pixel belongs to) within the neighborhood of each seed point. Unlike standard k-means, which searches the entire image, SLIC limits the search range to 2S*2S, which can accelerate algorithm convergence.

[0040] Step 2.4: Distance Measurement. This includes color distance and spatial distance. For each found pixel, calculate its distance to the seed point.

[0041] Step 2.5: Iterative optimization. Theoretically, the above steps are iterated until the error converges (which can be understood as the cluster center of each pixel no longer changing);

[0042] Step 2.6: Enhance Connectivity. The iterative optimization described above may result in the following drawbacks: multi-connectivity issues, excessively small superpixel sizes, and single superpixels being fragmented into multiple discontinuous superpixels. These issues can be addressed by enhancing connectivity. The main idea is to create a new label table where all elements are -1. Following a "Z" shaped approach (from left to right, from top to bottom), discontinuous superpixels and excessively small superpixels are reassigned to neighboring superpixels. Pixels that have been traversed are assigned their corresponding labels until all points have been traversed.

[0043] To verify the effectiveness of the method of the present invention, a simulation experiment was conducted based on the above technical solution. The specific results and analysis are as follows:

[0044] 1. Experimental images and experimental setup

[0045] The hyperspectral remote sensing images used in the experiment were selected from the hyperspectral gallbladder cancer dataset. Table 1 shows the basic information of the three datasets. The image size is 1024×1280, and the spectral bands are 60. The experimental test environment consisted of an Intel Core i5-7200U CPU, a 64-bit Windows 10 operating system, and a Tesla T4 GPU. The simulation tool used was the Keras deep learning framework, with Python version 3.6.6.

[0046] Table 1 Dataset Characteristics

[0047]

[0048] 2. Experimental Procedure and Results

[0049] To highlight the superiority of our method, we will compare its quantitative classification results with four other typical methods: PCA, LDA, and LPP. To ensure the accuracy of the comparative experiments, we tested the performance of each dimensionality reduction method on hyperspectral gallbladder cancer images with a fixed training sample size.

[0050] Table 2. Classification results (%) of four dimensionality reduction methods on the gallbladder cancer dataset.

[0051]

[0052] As shown in Table 2, the classification accuracy of each algorithm gradually improves with the increase of training samples. An increase in training samples means richer category information, allowing the extracted features to better characterize the differences between different land features, thus improving classification accuracy. With the same number of training samples, the SML algorithm consistently achieves the highest classification accuracy. This is because the SML algorithm considers both global and local information when constructing the projection matrix and performs spatial-spectral reconstruction on the original data, reducing information redundancy and noise interference. This allows the data in the low-dimensional space to maintain its original local structure, making the extracted features more discriminative.

Claims

1. A method for micro-hyperspectral gallbladder cancer image dimensionality reduction, characterized in that, It comprises the following steps: S1: using PaCMAP technology to reduce the dimension of the spectral dimension of hyperspectral image, the spectral dimension of the original data set in the experiment is compressed to 3 dimensions using PaCMAP method, greatly reducing the data amount on the basis of retaining effective spectral information, so as to achieve the purpose of reducing the training time, and then filtering the above dimension reduction data through AWF filter, which plays the effect of smoothing pixels; S2: based on the set number of superpixel segmentation, using simple iterative linear clustering method to segment the above filtered data, obtaining several superpixel blocks; using kernel principal component analysis method to reduce the dimension of each superpixel block, obtaining low-dimensional subspace low-dimensional feature; S3: the training samples in these low-dimensional image blocks will be sent to the support vector machine model network model, and the model will be trained; S4: after the support vector machine is trained, the test sample is input into the network for classification of the test sample, and the final classification label is generated through the maximum voter; The process of extracting features by popular learning PaCMAP in step S1 is as follows: A1: find a certain number of short-range pixel pairs, medium-range pixel pairs and long-range pixel pairs in the graph respectively; A2: initialize the embedding value of the pixel pair using PCA; A3: reduce the function value of the loss function to the minimum in three stages, and obtain the final category value corresponding to the pixel.

2. The method according to claim 1, wherein, The process of simple iterative linear clustering in step S2 is as follows: B1: initialize seed points: evenly distribute seed points in the image according to the set number of superpixels; assuming that the total number of pixels in the picture is N, and the pre-segmentation is K superpixels of the same size, then the size of each superpixel is N / K, and the distance between adjacent seed points is approximately S=sqrt(N / K); B2: reselect seed points in the n*n neighborhood of the seed points, taking n=3; the specific method is: calculate the gradient value of all pixel points in the neighborhood, and move the seed point to the place with the smallest gradient in the neighborhood; B3: assign a class label to each pixel point in the neighborhood around each seed point; unlike the standard k-means, the search range of SLIC is limited to 2S*2S, which can accelerate the convergence of the algorithm; B4: distance measurement; including color distance and spatial distance; for each searched pixel point, calculate its distance from the seed point; B5: iterative optimization; in theory, the above steps are iterated until the error converges; B6: enhance connectivity; after the above iterative optimization, the following defects may occur: multiple connectivity, superpixel size is too small, a single superpixel is cut into multiple discontinuous superpixels, etc., which can be solved by enhancing connectivity; the main idea is: create a mark table, all elements in the table are-1, and reassign the discontinuous superpixels and superpixels with too small size to adjacent superpixels in the order from left to right and from top to bottom, and assign the traversed pixel points to the corresponding label until all points are traversed.

Images