Classification method for medical hyperspectral image based on removing bad bands

Abstract

This invention relates to the field of medical hyperspectral image processing technology, and discloses a medical hyperspectral image classification method based on removing undesirable bands. The method includes: converting a three-dimensional medical hyperspectral image into a two-dimensional image; calculating the matched filter weight and information entropy weight for each band of the two-dimensional image; obtaining a fusion weight based on the matched filter weight and information entropy weight; sorting all bands of the two-dimensional image based on the fusion weight and target contribution, and filtering out undesirable bands; and performing classification operations on the two-dimensional image after removing undesirable bands. Based on the principles of matched filters and information entropy, the method calculates the average value of the absolutely normalized matched filter weights, and simultaneously calculates the average weight of the information entropy. The two weights are fused to evaluate each band for filtering, thus reducing the computational resource requirements and improving computational efficiency while ensuring classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical hyperspectral image processing technology, specifically to a medical hyperspectral image classification method based on removing undesirable bands. Background Technology

[0002] Hyperspectral images (HSI) have higher spectral resolution than traditional color digital images, typically containing dozens or even hundreds of bands. This rich spectral information can provide a basis for accurate target identification.

[0003] For Medical Hyperspectral Image (MHSI), the dramatic increase in dimensionality drastically increases the number of training samples required for parameter estimation. If the number of training samples is too small to meet the increased dimensionality requirements of the feature space, the accuracy of the estimated parameters will be affected, leading to unsatisfactory classification results. Although MHSI provides a wealth of more detailed biological tissue information, in some practical applications, the increased data volume does not necessarily increase the original information content. MHSI redundancy includes spatial and spectral redundancy. In a band image, the gray levels of sampling points on the same target surface typically exhibit spatial continuity, but gray levels represented by discrete pixel sampling do not fully utilize this feature, resulting in spatial redundancy. Due to MHSI's high spectral resolution and high data dimensionality, information in one band of the image can be partially or completely predicted from other bands, resulting in spectral redundancy. Large data volumes and computational demands, if not handled properly, can affect classification accuracy. Therefore, band selection (BS) is necessary to remove unwanted bands. For most HSI processing, removing undesirable bands, such as low signal-to-noise ratio (SNR) bands, is a necessary preprocessing step.

[0004] Currently, the most common methods for removing unwanted bands in MHSI are visual inspection and sensor setup. The former is very time-consuming, while the latter easily removes bands of acceptable quality or ignores some noisy bands. Based on wavelength information, noisy bands that cover water absorption characteristics are relatively easy to detect. However, unless every band in the image is traversed, it is difficult to find bands outside this wavelength range. Since hyperspectral images typically have more than 100 bands, the traversal process is very time-consuming.

[0005] Therefore, how to quickly and accurately remove unwanted bands, thereby reducing the computational load of MHSI image classification and improving classification efficiency, has become an urgent problem to be solved. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a medical hyperspectral image classification method based on removing undesirable bands, in order to solve the problem that the existing methods for removing undesirable bands in medical hyperspectral images have a long traversal process, resulting in low classification algorithm efficiency.

[0007] This invention provides a medical hyperspectral image classification method based on removing undesirable bands, comprising:

[0008] Converting three-dimensional medical hyperspectral images into two-dimensional images;

[0009] Calculate the matched filter weights and information entropy weights for each band of the two-dimensional image;

[0010] The fusion weights are obtained based on the matched filter weights and the information entropy weights.

[0011] Based on the fusion weight and target contribution, all bands of the two-dimensional image are sorted and defective bands are screened out.

[0012] After removing unwanted bands, classification operations are performed on the two-dimensional image.

[0013] Optionally, after converting the three-dimensional medical hyperspectral image into a two-dimensional image, the method further includes:

[0014] Band-by-band normalization processing of two-dimensional images.

[0015] Optionally, the matched filter weights for each band of the two-dimensional image are calculated, including:

[0016] Obtain the covariance of the two-dimensional data matrix of the two-dimensional image;

[0017] Based on the covariance of the two-dimensional data matrix of the two-dimensional image, the normalized weight vector corresponding to each target pixel in the current band is obtained;

[0018] Calculate the average value of the absolutely normalized matched filter weights for all target pixels; where the average value of the absolutely normalized matched filter weights is used as the matched filter weight for the current band.

[0019] Optionally, the information entropy weight of each band of the two-dimensional image is calculated, including:

[0020] Based on the entropy value and information entropy redundancy of the current band, obtain the information entropy weight corresponding to the current band;

[0021] The average of the information entropy weights corresponding to all bands is used as the information entropy weight for each band.

[0022] Optionally, the fusion weights are obtained based on the matched filter weights and the information entropy weights, including:

[0023] Set the fusion weight ratio of the matched filter weights to k;

[0024] Set the fusion weight ratio of the information entropy weight to 1-k;

[0025] Wherein, k ranges from 0.85 to 0.99.

[0026] Optionally, based on the fusion weights and target contribution, all bands of the two-dimensional image are sorted to filter out undesirable bands, including:

[0027] The target contribution of each band is evaluated based on the fusion weight, and then sorted in order from low to high.

[0028] Based on the preset number of undesirable bands, the bands ranked first are designated as undesirable bands.

[0029] Optionally, before performing classification operations on the two-dimensional image, the following steps are also included:

[0030] Accuracy assessment of two-dimensional images with defective bands removed was performed using support vector machines and test samples;

[0031] If the accuracy does not exceed the preset value, the fusion weights will be reset.

[0032] Optionally, any one of the following algorithms can be used to classify the two-dimensional image: support vector machine, K-nearest neighbor, decision tree, or Naive Bayes.

[0033] Beneficial effects of the embodiments of the present invention:

[0034] This invention addresses the problems of undesirable bands in medical hyperspectral images due to underutilization of spatial information and the poor accuracy of traditional bad band removal methods. Based on the principles of matched filters and information entropy, it selects multiple target pixels and calculates their corresponding normalized MF weight vector values, obtaining the average absolute normalized MF weight. Simultaneously, it calculates the average weight of information entropy and fuses these two weights to evaluate each band for screening. The intersection of low-contribution bands of all targets is considered an undesirable band, and the number of undesirable bands to be removed (Lb) is manually set. This medical hyperspectral image classification method based on undesirable band removal, provided by this invention, removes undesirable bands by fusing matched filters and information entropy before image classification, improving the accuracy of bad band removal and achieving better screening results than traditional methods. Image classification is then performed on this basis, ensuring classification accuracy while reducing computational resource requirements and improving computational efficiency. Attached Figure Description

[0035] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0036] Figure 1 A flowchart of a medical hyperspectral image classification method based on removing undesirable bands is shown in an embodiment of the present invention;

[0037] Figure 2 A flowchart of a method for removing bad bands based on a fused matched filter and an information entropy algorithm is shown in an embodiment of the present invention.

[0038] Figure 3 A flowchart of another medical hyperspectral image classification method based on removing undesirable bands is shown in an embodiment of the present invention;

[0039] Figure 4 An example of raw data image is shown in an embodiment of the present invention;

[0040] Figure 5 An example of a real image is shown in an embodiment of the present invention;

[0041] Figure 6 This shows an image after removing unwanted bands using a matched filter method;

[0042] Figure 7 An image is shown after removing unwanted bands using the information entropy method;

[0043] Figure 8 The image shown is an example of a medical hyperspectral image classification method based on removing undesirable bands, as described in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] This invention provides a medical hyperspectral image classification method based on removing undesirable bands, such as... Figure 1 As shown, it includes:

[0046] Step S10: Convert the three-dimensional medical hyperspectral image into a two-dimensional image.

[0047] In this embodiment, the three-dimensional medical hyperspectral image observation set X = [L, W, H] to be processed is imported, where L, W, and H represent the three dimensions of the data, respectively.

[0048] In a specific embodiment, the two-dimensional image is normalized band by band. The dataset is divided into an image consisting of H bands, and X = [L, W, H] is transformed into X' = [L*W, H]. The L*W bands are then normalized.

[0049] In a specific implementation, linear normalization is performed on each spectral band in the image. For the i-th band B in the image... i The normalization process is as follows:

[0050]

[0051] Wherein, min(B i ) and max(B i ) are respectively the i-th band B i The minimum and maximum spectral values ​​are calculated. Normalization is performed on all H bands so that the pixel values ​​of each band are between (0, 1), thus completing the normalization preprocessing.

[0052] Step S20: Calculate the matched filter weight and information entropy weight for each band of the two-dimensional image.

[0053] In this embodiment, calculating the matched filter (MF) weights for each band of the two-dimensional image includes:

[0054] Obtain the covariance of the two-dimensional data matrix of the two-dimensional image.

[0055] The normalized weight vector corresponding to each target pixel in the current band is obtained from the covariance of the two-dimensional data matrix based on the two-dimensional image.

[0056] Calculate the average value of the absolutely normalized matched filter weights for all target pixels; where the average value of the absolutely normalized matched filter weights is used as the matched filter weight for the current band.

[0057] In a specific embodiment, the covariance of the two-dimensional data matrix of a two-dimensional image is calculated as follows:

[0058]

[0059] Calculate M target pixels d (1) d (2) , ..., d (M) The corresponding M normalized MF weight vectors:

[0060] |w ～(1) |,|w ～(2) |,...,|w ～(M) |:

[0061] Among them, w ～(i) =[w～(i) 1, w ～(i) 2, ..., w ～(i) L ] T , 1≤i≤M.

[0062] For M target pixels d (1) d (2) , ..., d (M) Find the average of the absolutely normalized MF weights:

[0063]

[0064] Among them, |w~| avg =[|w~1| avg ,|w~2| avg , ..., |w~ L | avg ] T Let w mf =|w~| avg The average weight of the band selection by the MF method is obtained.

[0065] The calculation of the information entropy (IE) weights for each band of a two-dimensional image includes:

[0066] Based on the entropy value and information entropy redundancy of the current band, obtain the information entropy weight corresponding to the current band.

[0067] The average of the information entropy weights corresponding to all bands is used as the information entropy weight for each band.

[0068] In this embodiment, the information entropy weights of each band are calculated by uniformly selecting pixels.

[0069] Calculate the entropy value e for each band. j The calculation method is as follows:

[0070]

[0071] Where, p j This represents the percentage of the index corresponding to the j-th band among all possible values. j is used to distinguish different bands, and there are a total of m bands.

[0072] Calculate the weight w of band j j The calculation method is as follows:

[0073]

[0074] Where dj = 1 - e j , which represents the redundancy of information entropy.

[0075] The average weight w is calculated from the weights of each band. ie :

[0076]

[0077] w ie =|w|avg.

[0078] Step S30: Obtain the fusion weights based on the matched filter weights and the information entropy weights.

[0079] In this embodiment, the fusion weight ratio of the matched filter weights is set to k; the fusion weight ratio of the information entropy weights is set to 1-k; where k ranges from 0.85 to 0.99. In a specific embodiment, it has been verified that the data accuracy is optimal when k is 0.95: w′=0.95*w mf +0.05*w ie Where w′ represents the weights after fusion, w mf The weights obtained from the MF algorithm, w ie The weights are obtained from the IE algorithm. Each band is evaluated using the fused average weight, and the intersection of low-contribution bands of all targets can be considered as undesirable bands.

[0080] Step S40: Based on the fusion weight and target contribution, sort all bands of the two-dimensional image and filter out bad bands.

[0081] In this embodiment, the target contribution of each band is evaluated according to the fusion weight, and they are sorted in ascending order. Based on a preset number of bad bands, the bands ranked higher are designated as bad bands. In a specific embodiment, the number of bad bands Lb to be removed from the image is preset. The number of bad bands Lb is set according to the actual image requirements. The flowchart of the bad band removal method based on fusion matched filter and information entropy algorithm is as follows... Figure 2 As shown.

[0082] Step S50: After removing unwanted bands, perform classification operations on the two-dimensional image.

[0083] The result of MF target detection is a linear combination of all bands. The weight coefficient of each band, i.e., the linear combination of the weights obtained by the MF algorithm, can be considered as the contribution of each band to the detection of the target of interest. Information entropy is a noise-insensitive standard used to measure the hidden information in a random variable. The IE algorithm analyzes the contribution of each band to target detection by calculating the information entropy weights of each band. Under different conditions, the band selection effects of the two algorithms have their own advantages and disadvantages. Based on this, the method of fusing IE and MF can effectively extract and remove damaged bands.

[0084] This invention addresses the problems of undesirable bands in medical hyperspectral images due to underutilization of spatial information and the poor accuracy of traditional bad band removal methods. Based on the principles of matched filters and information entropy, it selects multiple target pixels and calculates their corresponding normalized MF weight vector values, obtaining the average absolute normalized MF weight. Simultaneously, it calculates the average weight of information entropy and fuses these two weights to evaluate each band for screening. The intersection of low-contribution bands of all targets is considered an undesirable band, and the number of undesirable bands to be removed (Lb) is manually set. This medical hyperspectral image classification method based on undesirable band removal, provided by this invention, removes undesirable bands by fusing matched filters and information entropy before image classification, improving the accuracy of bad band removal and achieving better screening results than traditional methods. Image classification is then performed on this basis, ensuring classification accuracy while reducing computational resource requirements and improving computational efficiency.

[0085] As an optional implementation, before performing classification operations on the two-dimensional image, the following steps are also included:

[0086] The accuracy of two-dimensional images with defective bands removed is evaluated using support vector machines and test samples.

[0087] If the accuracy does not exceed the preset value, the fusion weights will be reset.

[0088] In this embodiment, as Figure 3 As shown, the accuracy of the results after removing bad bands is evaluated using test samples. The remaining labeled samples used for training are used for testing, and the confusion matrix is ​​calculated to obtain the overall classification accuracy (OA) and Kappa coefficient. The classification accuracy and standard deviation are recorded. In a specific embodiment, the preset value for accuracy evaluation is set to 98%.

[0089] As an optional implementation, any one of the following algorithms can be used to classify the two-dimensional image: support vector machine, K-nearest neighbor, decision tree, or Naive Bayes.

[0090] This invention provides an automatic pre-removal method for bad bands by fusing IE and MF, applicable to the removal of undesirable bands in the In-Vivo human brain hyperspectral image dataset. The image contains 826 bands; 548 noisy bands are removed, leaving 278 bands. The image size is 443×479, with a spatial resolution of 128.7μm. The ground truth map mainly contains three tissue types and one background type.

[0091] Table 1 Comparison of Classification Accuracy and Computation Time

[0092]

[0093] As shown in Table 1, Figures 4-8As shown, the highest accuracy is achieved when the number of retained bands (band_num) is 278. Compared to the single MF algorithm and IE algorithm, the method provided in this embodiment of the invention can achieve higher screening accuracy.

[0094] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A medical hyperspectral image classification method based on removing bad bands, characterized in that, include: Converting three-dimensional medical hyperspectral images into two-dimensional images; Calculate the matched filter weights and information entropy weights for each band of the two-dimensional image; The fusion weights are obtained based on the matched filter weights and the information entropy weights. Based on the fusion weights and target contribution, all bands of the two-dimensional image are sorted and defective bands are screened out. After removing the unwanted bands, a classification operation is performed on the two-dimensional image; The calculation of the matched filter weights for each band of the two-dimensional image includes: Obtain the covariance of the two-dimensional data matrix of the two-dimensional image; Based on the covariance of the two-dimensional data matrix of the two-dimensional image, the normalized weight vector corresponding to each target pixel in the current band is obtained; Calculate the average value of the absolutely normalized matched filter weights for all the target pixels; wherein the average value of the absolutely normalized matched filter weights is used as the matched filter weight of the current band. Calculating the information entropy weight of each band of the two-dimensional image includes: Based on the entropy value and information entropy redundancy of the current band, obtain the information entropy weight corresponding to the current band; The average of the information entropy weights corresponding to all bands is used as the information entropy weight for each band. 2.The method of claim 1, wherein, After converting three-dimensional medical hyperspectral images into two-dimensional images, the process also includes: The two-dimensional image is normalized band by band. 3.The method of claim 1, wherein, The fusion weights are obtained based on the matched filter weights and the information entropy weights, including: The fusion weight proportion of the matched filter weights is set as k ; The fusion weight proportion of the information entropy weight is set to 1 k ; in, k The range is 0.85 to 0.

99.

4. The medical hyperspectral image classification method based on removing undesirable bands according to claim 1, characterized in that, Based on the fusion weights and target contribution, all bands of the two-dimensional image are sorted, and undesirable bands are filtered out, including: The target contribution of each band is evaluated according to the fusion weight, and sorted in order from low to high; Based on the preset number of undesirable bands, the bands ranked first are designated as the undesirable bands.

5. The medical hyperspectral image classification method based on removing undesirable bands according to claim 1, characterized in that, Before performing classification operations on the two-dimensional image, the following steps are also included: The accuracy of the two-dimensional image with defective bands removed is evaluated using a support vector machine and test samples. If the accuracy does not exceed the preset value, the fusion weights are reset.

6. The medical hyperspectral image classification method based on removing undesirable bands according to claim 1, characterized in that, The two-dimensional image is classified using any one of the following algorithms: Support Vector Machine, K-Nearest Neighbor, Decision Tree, or Naive Bayes.

Images