A Multimodal Medical Image Fusion Method Based on Frequency-Domain Fusion

Abstract

The present invention discloses a multi-modal medical image fusion method based on frequency-domain fusion, belonging to the technical field of medical image fusion, and comprising the following steps: S1: source image decomposition; S2: frequency-domain fusion; S3: denoising processing; S4: reconstruction processing. The present invention obtains different structural and texture detail information of the source image through decomposition. Compared with the discrete wavelet transform, the decomposed image greatly reduces the appearance of noise; the fusion method for the high-frequency part can effectively extract the structural and detail information of the image and reduce the appearance of noise. The fusion method for the structural texture part effectively extracts the image features in combination with the characteristics of deep learning. Finally, the Gaussian smoothing operation adopted can reduce the appearance of noise in this part; the fused image combines the features of different modalities, which is beneficial to the clinical diagnosis and treatment of doctors.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image fusion, and particularly relates to a multi-modal medical image fusion method based on frequency-domain fusion. Background Art

[0002] In modern medical diagnosis and treatment, medical image fusion plays an irreplaceable role. For some medical image modalities, such as magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomography (PET), etc., a single modality makes it difficult for doctors to make accurate judgments. Multi-modal medical image fusion integrates the functions of each modality to obtain a fused image of two or more modality medical images, which is beneficial to expert diagnosis and patient treatment.

[0003] Medical image fusion involves a wide range of general fusion techniques to integrate complementary information from different modalities. For medical analysis, image fusion provides great diversity of image features, usually resulting in good medical diagnoses. Additional information obtained from the fused image can be well utilized to accurately detect the lesion location. Due to the high requirements of multi-modal medical images, some fusion techniques have been developed in recent years. Generally, image fusion techniques can be classified into two categories, such as spatial domain and transform domain methods. Due to the characteristics of spatial information, spatial domain methods select pixels or blocks from each source image to construct a fused image, and the detailed information of the source image can be accurately preserved. All the information is preserved in the fused image as it is in the source image. Since spatial domain methods can perfectly preserve the spatial information of the source image, they have good performance in multi-focus and multi-exposure image fusion. However, the disadvantages of spatial domain methods are also significant. Spatial domain methods are difficult to integrate information from the same position of each source image, except for the weighted average method of image pixels, but the weighted average method of image pixels often causes a decrease in the contrast and clarity of the fused image, and this disadvantage is unacceptable in medical image fusion.

[0004] Different from the spatial domain method, the transform domain method is commonly used. The transform domain method first converts the source image into special coefficients; then, fuses the coefficients; finally, obtains a fused image through the inverse transform of all the coefficients. In recent years, the methods based on sparse representation and multi-scale transform have become the most popular fusion strategies. The method based on multi-scale transform is a frequency domain method in image processing. This algorithm decomposes the source image into high-frequency and low-frequency components through transforms (such as CVT, NSCT, NSST), and then designs different fusion rules for high-frequency and low-frequency components to obtain the fused image. The method based on sparse representation is a time domain processing method of images, which obtains the sparse representation of the original image through dictionary learning. Sparse representation and multi-scale transform decompose the source image in the time domain and frequency domain respectively, and are the two most mainstream image fusion methods in the transform domain. However, the computational cost based on sparse representation is usually relatively long. Therefore, the multi-scale transform method has become the most popular algorithm in transform domain-based image fusion. In multimodal image fusion, the multi-scale transform method is commonly used. Compared with the sparse representation method, the multi-scale transform method runs faster. However, for the multi-scale transform method, simple fusion rules cannot always successfully distinguish the details and structural information from the coefficients, which may cause a decrease in image quality. Therefore, it is very important to design an effective fusion rule.

[0005] In order to better extract the details of medical images and preserve more image energy, while also improving the visual perception of the images. For this purpose, a multimodal medical image fusion method based on frequency division domain fusion is proposed. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: how to better extract the details of medical images and preserve more image energy, while also improving the visual perception of the images. A multimodal medical image fusion method based on frequency division domain fusion is provided, which can decompose the source image into different frequency domains, adopt appropriate fusion methods in different frequency domains, and improve the retention of source image features; at the same time, make the structural and texture information of the source MRI and CT images more abundant in the fused image, play an auxiliary role for doctors in clinical treatment, and improve the diagnosis efficiency.

[0007] The present invention solves the above technical problems through the following technical solutions. The present invention includes the following steps:

[0008] S1: Decomposition of source images

[0009] Input the registered source image A and source image B into a weighted least squares filter for two-layer decomposition to obtain a high-frequency part and a low-frequency part. For the low-frequency part, perform structural texture decomposition through indirect gradient filtering, and finally obtain a three-layer decomposed image, namely the high-frequency part, the texture part, and the structural part;

[0010] S2: Frequency-domain fusion

[0011] Perform frequency-domain fusion on the images of each part of the decomposed source image A and source image B, and then obtain a high-frequency fusion image, a structure fusion image, and a texture fusion image;

[0012] S3: Denoising processing

[0013] Perform Gaussian smoothing on the obtained structure fusion image and texture fusion image to obtain a denoised feature image;

[0014] S4: Reconstruction processing

[0015] Adopt the inverse transformation processing of weighted least squares filtering and indirect gradient filtering, that is, add the high-frequency fusion image, the structure fusion image, and the texture fusion image after denoising the feature image to obtain the final fusion image.

[0016] Furthermore, in the step S1, the source image A and the source image B are respectively the source MRI image and the source CT image of the same part of the brain of the same patient.

[0017] Furthermore, in the step S1, it specifically includes the following process:

[0018] S11: Use a weighted least squares filter to decompose the i-th source image I i into two parts, namely the low-frequency part and the high-frequency part

[0019] S12: Put the low-frequency part into an indirect gradient filter, and use the following formula to re-adjust the gradient of the low-frequency part image and the corresponding indirect gradient:

[0020]

[0021] where π is the adjacent pixel of the central pixel p, I is the discrete signal, is the re-adjusted gradient, and W p is the re-adjusted weight, defined as follows:

[0022]

[0023] where ε is a constant;

[0024] S13: In order to obtain the filtering result of the low-frequency part , first define a temporary signal R. For each temporary signal R, the 1D guided filtering process finds the best linear transformation coefficients a p and b p , and minimize it, as described in the following formula:

[0025]

[0026] Among them, β is the smoothing parameter, σ is the scale parameter, and W n is the Gaussian weight, which is defined as follows:

[0027]

[0028] The structural part is obtained by the following formula:

[0029] S p = g σ (a p )R p + g σ (b p )

[0030] The texture part is obtained by the following formula:

[0031]

[0032] Furthermore, in the step S2, for the high-frequency part, phase consistency, local sharpness change, and local energy are introduced to jointly adjust the information of the fused image, and a high-frequency fused image is obtained.

[0033] Furthermore, in the step S2, the specific process of obtaining the high-frequency fused image is as follows:

[0034] S201: Regarding the image as a two-dimensional signal, the phase consistency of the image at the (x, y) position is calculated by the following formula:

[0035]

[0036] Among them, θ k is the direction angle at k, and θ k respectively represent the amplitude and angle of the k-th Fourier component; the parameter ε is a constant used to remove the DC component in the image signal, which is calculated by the following formula:

[0037]

[0038] Among them, and are the convolution results of the input image at (x, y);

[0039] S202: Introduce sharpness change, and the calculation formula for sharpness change is as follows:

[0040]

[0041] Among them, Ω represents a local area of size 3×3 input at (x, y), and (x0, y0) represents a pixel point in the local area Ω;

[0042] Meanwhile, calculate the neighborhood coefficient at (x, y), and the calculation formula for the local sharpness change is as follows:

[0043]

[0044] S203: Introduce the local energy, and the calculation formula is as follows:

[0045]

[0046] S204: Finally, the NAM algorithm formula is as follows:

[0047] NAM(x, y) = (PC(x, y)) α1 ·(LSCM(x, y)) β1 ·(LE(x, y)) γ1

[0048] Among them, α1, β1, and γ1 are parameters used to adjust PC, LSCM, and LE in NAM;

[0049] When NAM is obtained, the fused high-pass subband image is calculated according to the rules proposed by the following formula:

[0050]

[0051] Among them, H F 、H A and H B are the high-frequency subband fused image, source images A and B respectively;

[0052] S205: The final high-frequency fused image is obtained from the following formula:

[0053]

[0054] Among them, M×N represents the size of the sliding window centered at (x, y).

[0055] Furthermore, in the step S2, for the structure and texture parts, a convolutional neural network is used to extract depth features, and through a Softmax activation function, the weight maps of the structure and texture parts are obtained respectively, and then the structure fusion map and the texture fusion map are obtained based on the weight maps.

[0056] Furthermore, in the step S2, the specific process of obtaining the structure fusion map and the texture fusion map is as follows:

[0057] S211: Extract deep features using the pre-trained VGG19 convolutional neural network. For each layer of feature maps of each source image, it is defined as follows:

[0058]

[0059] Among them, I i represents the i-th source image, and is defined as the l-th layer feature map of the i-th source image. max() represents the ReLU operation. For each feature map, after l1 regularization calculation, we get as the l-th layer feature map, as shown in the following formula:

[0060]

[0061] The corresponding weight map is as follows:

[0062]

[0063] Among them, e () is the exponentiation based on e;

[0064] S212: Based on the obtained weight map, the result of the l-th layer image fusion is:

[0065]

[0066] Select the maximum pixel value of each layer, and the obtained fusion map is:

[0067]

[0068] Furthermore, in the step S3, the formula for Gaussian smoothing of the obtained structure fusion map and texture fusion map is as follows:

[0069]

[0070] Among them, w and h are the spatial dimensions of the weight map.

[0071] Furthermore, in the step S4, the final fusion image can be generated by the following formula:

[0072] F = F h + F l,s + F l,t .

[0073] The present invention has the following advantages compared with the prior art: The multi-modal medical image fusion method based on frequency division domain fusion decomposes to obtain different structural and texture detail information of the source images. The decomposed images greatly reduce the appearance of noise compared with the discrete wavelet transform; the fusion method for the high-frequency part can effectively extract the structural and detail information of the images and reduce the appearance of noise. The fusion method for the structural texture part effectively extracts the image features in combination with the characteristics of deep learning. Finally, the Gaussian smoothing operation adopted can reduce the appearance of noise in this part; the fused image combines the features of different modalities, which is beneficial to the clinical diagnosis and treatment by doctors. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a schematic flowchart of the multi-modal medical image fusion method based on frequency division domain fusion in Embodiment 1 of the present invention;

[0075] Figure 2 is a schematic flowchart of the implementation process of the detailed steps in Embodiment 1 of the present invention;

[0076] FIG. 3(a) is the source CT image in Embodiment 1 of the present invention;

[0077] FIG. 3(b) is the longitudinal edge feature image obtained by performing discrete wavelet transform on the source CT image in Embodiment 1 of the present invention;

[0078] FIG. 3(c) is the transverse edge feature image obtained by performing discrete wavelet transform on the source CT image in Embodiment 1 of the present invention;

[0079] FIG. 3(d) is the diagonal feature image obtained by performing discrete wavelet transform on the source CT image in Embodiment 1 of the present invention;

[0080] Figure 4 is the decomposition strategy schematic diagram in Embodiment 1 of the present invention;

[0081] Figure 5 is a schematic flowchart of the multi-modal medical image fusion method based on frequency division domain fusion in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0082] The following provides a detailed description of the embodiments of the present invention. These embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0083] Embodiment 1

[0084] As Figure 1 shown, this embodiment provides a technical solution: A multi-modal medical image fusion method based on frequency division domain fusion, including the following main steps:

[0085] Decomposition step: Input the registered source MRI image (source image A) and source CT image (source image B) into a weighted least squares filter for two-layer decomposition to obtain a high-frequency part and a low-frequency part. Since the low-frequency part is similar to the source image, perform structure-texture decomposition on the low-frequency part through indirect gradient filtering. Finally, obtain a three-layer decomposed image, namely: the high-frequency part, the texture part, and the structure part.

[0086] Fusion step: Perform fusion on each part of the decomposed MRI and CT images. For the noise in the high-frequency part, introduce phase consistency, local sharpness change, and local energy to jointly adjust the information of the fused image to obtain a high-frequency fused image. For the structure and texture parts, use a convolutional neural network to extract deep features, and through a Softmax activation function, obtain the weight maps of the structure and texture parts respectively. Then, based on the weight maps, obtain the structure fused image and the texture fused image.

[0087] Denoising step: To eliminate unnecessary noise generation, perform Gaussian smoothing on the obtained structure fused image and texture fused image to obtain a denoised feature image.

[0088] Reconstruction step: Perform inverse transformation processing using weighted least squares filtering and indirect gradient filtering, that is, add the denoised feature images of the high-frequency fused image, the structure fused image, and the texture fused image to obtain the final fused image.

[0089] The detailed steps are as follows (see Figure 2 ):

[0090] I. Decomposition step

[0091] As shown in Figure 3(b) , 3(c) , 3(d), the source CT image in Fig. 3(a) is decomposed into images in three directions through discrete wavelet transform, namely: the longitudinal edge feature image, the transverse edge feature image, and the diagonal feature image. It can be seen that there is obvious noise in the image after discrete wavelet transform. Therefore, the following decomposition strategy (as shown in Figure 4 ) is adopted in this embodiment, which can effectively reduce the appearance of noise.

[0092] 1). Use a weighted least squares filter to decompose the i-th source image I i into two parts, namely: the low-frequency part and the high-frequency part . As can be seen from Figure 4 , the noise in the decomposed image is significantly reduced compared to the image decomposed by discrete wavelet transform.

[0093] 2). As can be seen from Figure 4 , the low-frequency part is similar to the input source image Ii , therefore, put the low-frequency part of the image into the indirect gradient filter, and use the following formula to re-adjust the gradient of the low-frequency part of the image and the corresponding indirect gradient:

[0094]

[0095] where π is the adjacent pixel of the central pixel p, I is the discrete signal, is the re-adjusted gradient, and W p is the re-adjusted weight, which is defined as follows:

[0096]

[0097] where ε is a constant. The larger the value of ε, the more sensitive the algorithm is to noise. However, if ε is too large, the texture cannot be completely filtered. Therefore, in this embodiment, ε is set to 10 -4 .

[0098] To obtain the filtering result of the low-frequency part , first define a temporary signal R. For each temporary signal R, the 1D guided filtering process finds the best linear transformation coefficients a p and b p , and minimizes it as described by the following formula:

[0099]

[0100] where β is the smoothing parameter, σ is the scale parameter, and W n is the Gaussian weight, which is defined as follows:

[0101]

[0102] The structure part is obtained by the following formula:

[0103] S p = g σ (a p )R p + g σ (b p )

[0104] The texture part is obtained by the following formula:

[0105]

[0106] II. Fusion steps

[0107] 1). Fusion method of the structure and texture parts (V-CNN)

[0108] To better extract deep features, the pre-trained VGG19 convolutional neural network is used to extract more effective information. The feature maps of each layer for each source image are defined as follows:

[0109]

[0110] Among them, I i represents the i-th source image, and is defined as the l-th layer feature map of the i-th source image. max() represents the ReLU operation. For each feature map, after l1 regularization calculation, is obtained as the l-th layer feature map, as shown in the following formula:

[0111]

[0112] The corresponding weight map is as follows:

[0113]

[0114] Among them, e () is the exponentiation based on e;

[0115] To eliminate unnecessary noise generation, after obtaining the weight map, Gaussian smoothing is adopted where w and h are the spatial dimensions of the weight map.

[0116] Finally, the result of the l-th layer image fusion based on the obtained weight map is:

[0117]

[0118] To obtain the optimal fusion result for each layer, the maximum pixel value of each layer is selected, and the obtained fusion map is:

[0119]

[0120] 2) High-frequency part fusion method (NAM algorithm)

[0121] The key to high-pass sub-band fusion is to enhance the detailed features of each source image. Usually, lesions are often discovered through detailed information. To make the high-pass sub-band image contain more information, phase consistency is used to enhance the features of the image. Using phase consistency to process the image can improve the robustness of central coordinate extraction. In the high-pass sub-band, the value of phase consistency (PC) corresponds to the sharpness of the image object. Therefore, PC is used as the phase of the coefficient with the maximum local sharpness. Since the image can be regarded as a two-dimensional signal, the PC at the (x, y) position of the image can be calculated by the following formula:

[0122]

[0123] where θ k is the direction angle at k, and θ k represent the amplitude and angle of the k-th Fourier component respectively; the parameter ε is a constant used to remove the DC component in the image signal. In this embodiment, the parameter ε is set to 0.001. It can be calculated by the following formula:

[0124]

[0125] where and are the convolution results of the input image at (x, y);

[0126] As a contrast variable, PC cannot reflect local contrast changes. To make up for the deficiency of PC, sharpness change (SCM) is introduced, and its calculation formula is as follows:

[0127]

[0128] where Ω represents a local area of size 3×3 at (x, y). (x0, y0) represents a pixel point in the local area Ω; meanwhile, the neighborhood coefficient at (x, y) is calculated, and the calculation formula of local sharpness change (LSCM) is as follows:

[0129]

[0130] Since LSCM and PC cannot fully reflect local brightness information, local energy (LE) is introduced and can be calculated by the following formula:

[0131]

[0132] The final NAM algorithm formula is as follows:

[0133] NAM(x, y) = (PC(x, y)) α1 ·(LSCM(x, y)) β1 ·(LE(x, y)) γ1

[0134] where α1, β1, γ1 are parameters used to adjust PC, LSCM, and LE in NAM. In this embodiment, the parameters α1, β1, and γ1 are also set to 1, 2, and 2.

[0135] When NAM is obtained, the fused high-pass subband image can be calculated by the rule proposed by the following formula:

[0136]

[0137] where HF , H A and H B are the high-frequency sub-band fusion image, source images A and B, respectively;

[0138] The final high-frequency fusion image is obtained from the following formula:

[0139]

[0140] where M×N represents the size of the sliding window centered at (x, y).

[0141] III. Denoising Steps

[0142] The Gaussian smoothing operation is a two-dimensional convolution operation. To remove noise, Gaussian smoothing is used to perform a Gaussian smoothing operation on the image after structural texture decomposition. The formula is as follows:

[0143]

[0144] where w and h are the spatial dimensions of the weight map.

[0145] IV. Reconstruction Steps

[0146] The final fusion image can be generated by the following formula:

[0147] F = F h + F l,s + F l,t

[0148] Example 2

[0149] As Figure 5 shown, a multi-modal medical image fusion method based on frequency-domain fusion in this embodiment specifically includes the following main steps:

[0150] Obtain MRI and CT images of the same part of the same brain of the same patient (i.e., source images A and source images B);

[0151] Decompose the MRI image to obtain the first high-frequency image and the first low-frequency image, and decompose the CT image to obtain the second high-frequency image and the second low-frequency image;

[0152] Decompose the first low-frequency image and the second low-frequency image into structural and texture parts respectively;

[0153] According to the high-pass characteristics of the first high-frequency image and the second high-frequency image, use the NAM fusion method for fusion and denoising;

[0154] According to the structural and textural properties of the structural texture, use V-CNN to perform image fusion for the two modalities respectively;

[0155] The image after V-CNN fusion is denoised by Gaussian smoothing;

[0156] The high-frequency part of the fused image and the texture and structure part of the fused image are reconstructed to obtain the final fused image.

[0157] For the specific processing steps, reference may be made to Embodiment 1, which will not be elaborated here.

[0158] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-modal medical image fusion method based on frequency division domain fusion, characterized in that It includes the following steps: S1: Source image decomposition Input the registered source image A and source image B into a weighted least squares filter for two-layer decomposition to obtain a high-frequency part and a low-frequency part. For the low-frequency part, perform structure-texture decomposition through indirect gradient filtering, and finally obtain a three-layer decomposed image, namely the high-frequency part, the texture part, and the structure part; S2: Frequency-domain fusion Perform frequency-domain fusion on each part of the decomposed source image A and source image B to obtain a high-frequency fusion map, a structure fusion map, and a texture fusion map; In step S2, for the high-frequency part, introduce phase consistency, local sharpness change, and local energy to jointly adjust the information of the fused image to obtain a high-frequency fusion map; In step S2, the specific process of obtaining the high-frequency fusion map is as follows: S201: Regard the image as a two-dimensional signal, and calculate the phase consistency of the image at the (x, y) position through the following formula: where θ k is the direction angle at k, and θ k represent the amplitude and angle of the k-th Fourier component respectively; the parameter ε is a constant used to remove the DC component in the image signal, which is calculated by the following formula: Among them, and are the convolution results of the input image at (x, y); S202: Introduce sharpness change, and the calculation formula of sharpness change is as follows: Where, Ω represents a local area of size 3×3 at (x, y), and (x0, y0) represents a pixel point in the local area Ω; At the same time, calculate the neighborhood coefficient at (x, y), and the calculation formula of local sharpness change is as follows: S203: Introduce local energy, and the calculation formula is as follows: S204: Finally, the NAM algorithm formula is as follows: NAM(x,y) = (PC(x,y)) α1 ·(LSCM(x,y)) β1 ·(LE(x,y)) γ1 Where, α1, β1, γ1 are parameters used to adjust PC, LSCM, and LE in NAM; After NAM is obtained, the fused high-pass subband image is calculated according to the rule proposed by the following formula: Among them, H F , H A and H B are the high-frequency sub-band fused image, source images A and B respectively; S205: The final high-frequency fusion map is obtained from the following formula: Where, M×N represents the size of the sliding window centered at (x, y); In step S2, for the structure and texture parts, use a convolutional neural network to extract deep features, and through a Softmax activation function, obtain the weight maps of the structure and texture parts respectively, and then obtain the structure fusion map and the texture fusion map based on the weight maps; S3: Denoising processing Perform Gaussian smoothing on the obtained structure fusion map and texture fusion map to obtain a denoised feature image; S4: Reconstruction processing Adopt the inverse transformation processing of weighted least squares filtering and indirect gradient filtering, that is, add the high-frequency fusion map, the structure fusion map, and the texture fusion map after passing through the denoised feature image to obtain the final fused image.

2. The multimodal medical image fusion method based on frequency-domain fusion according to claim 1, wherein: In step S1, source image A and source image B are the source MRI image and source CT image of the same part of the brain of the same patient respectively.

3. A multi-modal medical image fusion method based on frequency-domain fusion according to claim 1, characterized in that: In step S1, specifically includes the following process: S11: Decompose the i-th source image I into two parts using a weighted least squares filter, namely the low-frequency part i and the high-frequency part respectively S12: Put the low-frequency part into an indirect gradient filter, and use the following formula to re-adjust the gradient of the low-frequency part image and the corresponding indirect gradient: where π is an adjacent pixel of the central pixel p, and I is a discrete signal, is the re-adjusted gradient, W p is the re-adjusted weight, which is defined as follows: Where, ε is a constant; S13: To obtain the filtering result of the low-frequency part first define a temporary signal R. For each temporary signal R, the 1D guided filtering process finds the best linear transformation coefficients a p and b p to minimize it, as described in the following formula: where β is the smoothing parameter, σ is the scale parameter, and W n is the Gaussian weight, defined as follows: The structure part is obtained from the following formula: S p = g σ (a p )R p + g σ (b p ) The texture part is obtained from the following formula:

4. A multi-modal medical image fusion method based on frequency-domain fusion according to claim 3, characterized in that: In step S2, the specific process of obtaining the structure fusion map and the texture fusion map is as follows: S211: Use the pre-trained VGG19 convolutional neural network to extract deep features, and define the feature map of each layer of each source image as follows: Among them, I i represents the i-th source image, and is defined as the l-th layer feature map of the i-th source image. max() represents the ReLU operation. For each feature map, after l1 regularization calculation, it is obtained as the l-th layer feature map, as shown in the following formula: The corresponding weight map is as follows: where e () is the exponentiation based on e; S212: The result of image fusion at the l-th layer is obtained based on the obtained weight map as follows: Select the maximum pixel value of each layer, and the obtained fused image is:

5. A multimodal medical image fusion method based on frequency division domain fusion according to claim 1, characterized in that: In the step S3, the formula for performing Gaussian smoothing on the obtained structural fusion map and texture fusion map is as follows: where w and h are the spatial dimensions of the weight map.

6. A multi-modal medical image fusion method based on frequency-domain fusion according to claim 1, characterized in that: In the step S4, the final fused image can be generated by the following formula: F = F h + F l,s + F l,t 。

Images