TWI806613B - Hyperspectral characteristic band selection method and active hyperspectral imaging device using the method - Google Patents

Abstract

The present invention discloses a hyperspectral feature band selection method and an active hyperspectral imaging device using the method. The method calculates the endmember number of the hyperspectral image of a predetermined disease with a virtual dimension algorithm, and uses the endmember number Between 1 and 2 times the predetermined number of selected bands, and then select the characteristic bands that match the aforementioned number of selected bands from most of the spectral bands in the hyperspectral image. The device mainly includes a base, a plurality of light sources and a sensing part, each of the light sources is respectively arranged on the base, and the wave bands emitted by them are respectively configured according to the characteristic wave bands calculated by the aforementioned method. The sensing part is arranged on the base part apart from the light sources, and is used for receiving the reflected light of an external object irradiated by the light sources to obtain a hyperspectral image.

Description

Hyperspectral characteristic band selection method and active hyperspectral imaging device using the method

The present invention is related to spectral band selection and image detection technology, in particular to a method for selecting bands using hyperspectral features and an active hyperspectral imaging device using the method.

In recent years, with the development of science and technology, image analysis and computing capabilities have been improved to a considerable extent, which can meet the needs of computing operations on large amounts of data. For example, hyperspectral imaging (HSI) has been promoted. Related research is booming. Among them, the hyperspectral imaging system refers to the collection of spatial information and spectral information of the object to be inspected by means of optical inspection, and integrated calculation. It has the characteristics of non-contact, no need to destroy the object to be inspected, and fast operation. To put it simply, the hyperspectral imaging system detects spectral bands that can cover many unknown material signals that cannot be seen by the human eye, that is, collects more than a hundred continuous spectral bands (spectral bands) information, and then stacks them into a hyperspectral image cube (hypercube).

For this reason, with the advantages of combining high spatial and spectral resolution, the hyperspectral imaging system is considered to be an analysis and detection tool with high development potential and prospects, and has been tried to be applied to industrial products, biomedical materials detection, and food safety of agricultural products. , burns and scalds and biomedicine and other industries. Among them, due to the non-invasive characteristics of the hyperspectral imaging system, and by shortening the object distance, spectral information of large areas and small details can be obtained. For medical detection such as disease diagnosis and surgical guidance, especially in the autonomous detection of early disease It has better application potential, making hyperspectral imaging system a medical application field worthy of further exploration.

For example, the journal article "Hyperspectral imaging for skin assessment in systemic sclerosis: a pilot study" published by Chen, Y.-M.Et al. in 2020 has preliminarily proved that hyperspectral imaging technology is effective in assessing scleroderma Feasibility of patient disease severity. However, the method used in this paper is a simple spectral angle mapper (SAM) method. The area analyzed by this method is limited to the ultrasonic imaging area (ROI), and 9 Each location contains 3*3 pixels, so there will be a total of 81 pixels. After averaging the spectral profiles of these 81 pixels, the disease is evaluated by spectral difference analysis severity. However, the spectral difference analysis method used in this paper can only compare the difference between two spectral waveforms, and is easily disturbed by other external environmental noises (such as skin hair, moles, or rupture of microvessels, etc.); and , the area analyzed in this paper is only a local area of interest (such as the part of the hand), so the correlation of the spectrum of the entire area of interest cannot be considered.

Moreover, in terms of structural design, traditional hyperspectral imaging instruments currently on the market mainly include spectroscopes (model Imspector N17E, SPECIM, Oulu, Finland), indium gallium arsenide microscopic microscopes (InGaAs, model Xeva-1.7-320 , Xenics, Leuven, Belgium) and standard halogen lamp (model 3900e DC, Illumination Technologies, Inc., New York, USA), where, since the light of the halogen lamp is full-band irradiation, the near-infrared light must first be filtered through the spectroscope (Wavebands between 900nm-1700nm) can only be used for spectral image collection. Moreover, in order to achieve a predetermined irradiation intensity, a ring-shaped halogen lamp is usually used, and its structural design needs to be configured with light guides, wires, and electronic control modules. In this way, the design of the beam splitter and the ring-shaped halogen lamp, It will increase the size of the instrument, which is obviously too bulky. If it is to be used clinically, it is necessary to reduce the weight of the instrument first, so as to have the opportunity to use it in the clinic. Physicians directly take pictures of the parts of patients that need to be analyzed, which can be used as an auxiliary tool for clinical diagnosis or evaluation of treatment effectiveness.

In addition, because there is a high correlation between the spectral bands contained in the hyperspectral image, and the redundancy between the spectral bands is relatively high, if no data screening is performed, the data volume It is still too large, takes up a lot of computing resources, and also takes a long time to compute. Moreover, for different diseases, it may only have analytical value in certain specific spectral bands. Accordingly, if the specific spectral band of a predetermined disease can be found, so as to simplify the amount of data, simplify the hyperspectral image analysis steps, and use it to reduce the components of traditional instruments and equipment, it will be considered by the relevant industry.

The main purpose of the present invention is to provide a method for selecting hyperspectral characteristic bands, which is mainly to select the characteristic bands of predetermined diseases under the premise of ensuring the accuracy of analysis, so as to simplify the amount of data, simplify the hyperspectral image analysis steps, and As a reference basis for reducing traditional instrument and equipment components.

The reason is, in order to achieve the above purpose, the present invention discloses a hyperspectral characteristic band selection method, which mainly calculates the endmember number of the hyperspectral image of a predetermined disease with a virtual dimension algorithm, and uses 1 times the endmember number to Between 2 times as the predetermined number of selected bands, the number is defined as n, and n is a natural number that is not zero, and then, from the majority of spectral bands in the hyperspectral image, select the number of bands corresponding to the aforementioned selected bands Matching number of characteristic bands.

In one embodiment, when the predetermined disease is scleroderma, the number of selected bands is preferably 1.5 times the number of endmembers; the predetermined disease is diabetes, and the number of selected wavebands is preferably 1 time or 1.5 times the number of endmembers.

In one embodiment, the band prioritization method is used to calculate the priority scores of all spectral bands, and the first n spectral bands with higher priority scores are used as characteristic spectra. Among them, the band prioritization method selects the variance, skewness, kurtosis, entropy or information divergence to calculate the spectral band priority score. Moreover, when performing operational analysis with the band prioritization method, it is better to select the number of bands that is 1.5 times the number of endmembers.

In one embodiment, the present invention further selects characteristic bands from the spectral bands by means of all band analysis or averaging. . Among them, when using the average method for calculation and analysis, it is better to select the number of bands to be 1 times the number of endmembers.

In one embodiment, the number of repetitions of those characteristic bands selected by all the band analysis methods, the average method and the band prioritization method are counted, and the number of repetitions reaching the predetermined threshold is taken as the better feature spectrum.

In another embodiment of the present invention, an active hyperspectral imaging device is disclosed, which is equipped with a light source corresponding to the characteristic band of a predetermined disease. Compared with the traditional hyperspectral imager, there is no need to reserve a filter , beam splitter and other components, and save the configuration of the catheter and wires of the ring halogen lamp, so as to achieve the purpose of lightweight, hand-held design, and as an evaluation of early detection or diagnosis of diseases.

Specifically, the active hyperspectral imaging device mainly includes a base, a plurality of light sources and a sensing part, wherein each of the light sources is respectively arranged on the base, and the wavelength bands emitted by each of the light sources are respectively based on the aforementioned The hyperspectral feature band selection method is used to configure the feature band or better feature band. The sensing part is arranged on the base part apart from the light sources, and is used for receiving the reflected light of an external object irradiated by the light sources to obtain a hyperspectral image.

In one embodiment, the light sources are centered on the position of the sensing part and arranged on the base in a circular arrangement.

In one embodiment, the base includes a shell, a cover, and a base, wherein the shell has a shell and an assembly port, and the assembly port is opened on the shell to make the inner space of the shell Connect with the outside world. The cover system has a cover body, a first hole and a recess, wherein the cover body is plate-shaped and connected to the shell body corresponding to the assembly port, and the first hole runs through the cover body At the central position, the recess is suitably recessed on one side of the cover body and communicates with the first hole.

The seat system has an annular seat body and a second hole, and the second hole is axially installed on the seat body along the center of the seat body, and the seat body is used to insert into the recess , and make the second hole coaxial with the first hole. Wherein, the light sources are respectively arranged on the base body in a ring shape with the second hole as the center.

In one embodiment, the sensing unit includes a main body, a cylindrical image capturing lens and a processing unit, wherein the main body is disposed in the inner space of the casing. The image-taking lens is arranged on the body, and is used to pass through the first hole and the second hole, so that one end of the column shaft of the lens is exposed outside the casing, and the other end of the column shaft of the lens is located at the The hyperspectral image is captured in the inner space of the casing. The processing unit is arranged on the main body and electrically connected with the image capturing lens for receiving and analyzing the hyperspectral image.

In one embodiment, the seat is made of ceramic material.

A: Active hyperspectral imaging device

10: base

11: Housing

111: shell body

112: Assembly port

113: Internal space

12: Cover body

121: cover body

122: The first hole

123: concave

13: seat body

131: seat body

132: Second hole

20: light source

30: Sensing part

31: Ontology

32: Image capture lens

33: Processing unit

40: frame body

Fig. 1 is a schematic flow chart of Example 1 of the present invention.

FIG. 2 is a schematic diagram of a general image of the sole of a subject in Example 1 of the present invention.

FIG. 3 is a schematic diagram of a hyperspectral image of a subject's plantar in Example 1 of the present invention.

Fig. 4 is a mask map of the sole of a subject in Example 1 of the present invention.

Fig. 5A to Fig. 5D are the distribution diagrams of the spectral bands selected by the third band selection method, the fourth band selection method, the fifth band selection method and the sixth band selection method in Example 1 of the present invention respectively, and are marked with pick respectively The spectral wavelength position of the outgoing band.

FIG. 6 is a graph of ROC curves of different band selection methods in Example 1 of the present invention.

Fig. 7 is a schematic flow chart of Example 2 of the present invention.

Fig. 8A to Fig. 8F are the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method and the sixth band selection method in the second example of the present invention The distribution map of the spectral bands, and the spectral wavelength positions of the selected bands are marked respectively.

Fig. 9A to Fig. 9G are the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method and the second band selection method in the second example of the present invention. Box plot of the statistical analysis of the calculation results of the seven-band selection method.

Fig. 9H is a box plot of statistical analysis by Mann-Whitney U test according to the prior art.

Fig. 10A to Fig. 10G are the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method and the second band selection method in the second example of the present invention. ROC curves between individual methods such as seven-band selection method and skin fraction, skin thickness and spectral difference analysis method.

FIG. 11 is a three-dimensional assembled view of the active hyperspectral imaging device of the present invention.

FIG. 12 is an exploded perspective view of the active hyperspectral imaging device of the present invention.

FIG. 13 is a schematic diagram of the position of the light source and its spectral band of the active hyperspectral imaging device of the present invention.

FIG. 14 is a schematic diagram of a specific embodiment of the present invention, which shows that the active hyperspectral imaging device is arranged on a frame.

15A to FIG. 15G are different characteristic wavebands (ie, 1100nm, 1150nm, 1200nm, 1300nm, 1450nm, 1550nm, and 1650nm) detection diagrams of light source intensity and uniformity, respectively.

Firstly, the specific nouns disclosed in the present invention are listed and described.

The "healthy person" disclosed in the present invention refers to a patient without a disease or a tissue sample thereof.

The "hyperspectral imaging information" disclosed in the present invention refers to a hyperspectral imaging device with a near-infrared light band (wavelength 900nm-1700nm) to shoot a predetermined target area to obtain a hyperspectral imaging (Hyperspectral Imaging), and then obtain the high The spectral image first uses the image augmentation technology to form the hyperspectral image into hyperspectral image cubes (Hyperspectral Image Cubes). Wherein, the target area can be but not limited to the skin of any part of the hands, arms, face, neck and feet, and can be classified as normal human skin tissue or abnormal human tissue based on the health status of the sampled human body.

The "hyperspectral waveform difference analysis method" disclosed in the present invention means that when establishing standard spectral waveform diagrams, disease spectral waveform diagrams, and normal spectral waveform diagrams, discrete images must first be excluded to improve the accuracy of the data. The algorithms used include Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and mean square error (MSE); in addition, the hyperspectral waveform The difference analysis method can also be used to calculate the difference between the spectral waveforms, so as to obtain the prevalence degree of the disease to be detected.

The so-called "spectral angle matching method" in the present invention refers to using the concept of n-dimensional angles to match the similarity between the spectral vector signal and the target spectral vector signal. The smaller the angle, the closer the target spectral vector signal is. Calculated as follows:

Wherein, L represents the number of spectral bands (numbers of bands).

The "spectral information divergence method" disclosed in the present invention is derived from the concept of divergence in information theory, and is used to measure the difference in probability behavior between two pixel vector spectral features. In other words, the spectral similarity between two pixel vectors is measured by the difference between the derived probability distributions based on their corresponding spectral features. Therefore, the difference between the spectral information divergence method and the spectral angle matching method is that the spectral angle matching method is to compare the angular geometric characteristics between two pixel vectors, while the spectral information divergence method is to measure the spectrum of two pixel vectors The distance between the probability distributions produced by the features. because Therefore, the spectral information divergence method can capture the variability between spectra more effectively than the spectral angle matching method.

The "image pre-processing technology" disclosed in the present invention is to remove non-interesting regions in the hyperspectral image, that is, to extract the predetermined target region in the image, and then eliminate discrete data through the hyperspectral waveform difference analysis method.

The "Hyperspectral Imaging Soft Abundance Scorer (HISAS) Algorithm" disclosed by the present invention includes the following steps: determination of Numbers of Endmembers, band selection, and minimum spectral decorrelation energy Chemical method, calculation of abundance score and abundance map.

The "endmember number" or "endmember value" disclosed in the present invention refers to the number of endmember components. This value is directly generated from the image. Currently, the commonly used algorithm is the virtual dimension (Virtual Dimensions, VD) algorithm. There are also Harsanyi-Farrand-Chang algorithm, Pixel Purity Index (PPI), algorithm with the largest simplex volume (N-FINDR), iterative error analysis (Iterative Error Analysis, IEA), vertex component analysis algorithm (Vertex Component Analysis (VCA), Minimum Volume Simplex Analysis (MVSA), Simplex Identification via Split Augmented Lagrangian (SISAL), Minimum Volume Enclosing Simplex Algorithm (MVES), Minimum Volume Constraint Nonnegative Matrix Factorization (MVCNMF) and Iterated Constrained Endmember (ICE).

，及協方差矩陣

，並分別求取兩個矩陣之特徵值與差值。其中，

表示樣本光譜向量集合，μ表示樣本光譜向量的平均光譜向量，

和

分別表示R _L×L 和K _L×L 兩個矩陣的特徵值，並稱為自相關特徵值和協方差特徵值，L表示為光譜波段數目。如果影像數據中存在一確定的高光譜信號源，則一定會存在某個光譜維度l，1

l

L，且若有

>λ _l 情況成立，因為信號影響樣本平均值，則會增加樣本自相關矩陣R _L×L 的方差，然而雜訊卻不會。此外，樣本的自相關矩陣R _L×L 與樣本的排列順序

無關。 The "virtual dimension algorithm" disclosed in the present invention is described as follows: assuming that the hyperspectral signal is an unknown and definite signal source, the signal spectrum only affects the first-order statistic (i.e., the sample average value), and the noise is the average value of Gaussian white noise of 0, and only exists in the second-order statistics of the sample. Next, calculate the autocorrelation matrix of the sample as

, and the covariance matrix

, and calculate the eigenvalues and differences of the two matrices respectively. in,

Represents the set of sample spectral vectors, μ represents the average spectral vector of sample spectral vectors,

and

Respectively represent the eigenvalues of the two matrices R _{L × L} and K _{L × L} , and are called autocorrelation eigenvalues and covariance eigenvalues, and L represents the number of spectral bands. If there is a definite hyperspectral signal source in the image data, there must be a certain spectral dimension l , 1

l

L , and if any

> λ _l is true because the signal affects the sample mean, which increases the variance of the sample autocorrelation matrix R _{L × L} , while the noise does not. In addition, the autocorrelation matrix R _{L × L} of the sample is related to the arrangement order of the sample

irrelevant.

Next, in order to determine the number of end members with large spectral differences, the correlation calculation is as follows:

versus

Among them, the null hypothesis (Null Hypothesis) H ₀ and the alternative hypothesis (Alternative Hypothesis) H ₁ correspond to the situation that the eigenvalue of the autocorrelation matrix is equal to the eigenvalue of the covariance matrix, and the eigenvalue of the autocorrelation matrix is greater than the eigenvalue of the covariance matrix. In other words, if the opposite hypothesis H ₁ is true (that is, the null hypothesis H ₀ is not true), it means that there is a definite hyperspectral signal, which affects the first-order statistics expressed by the eigenvalues of the autocorrelation matrix R _{L × L} , making its characteristic The value is greater than the eigenvalue of the covariance matrix K _{L × L} ; if the component only contains noise, the eigenvalue of its autocorrelation matrix R _{L × L} is equal to the eigenvalue of the covariance matrix K _{L × L.} That is, the signal spectrum will change the spectral mean value vector, which is non-zero in R _{L × L} , but is 0 in K _{L × L.}

，假設其確定的假警報率(false alarm rate)P_F為α，進行尋找下列公式(2-2)隨機決策規則中的閾值τ _l 之最大化偵測能力P_D。 Then, using the Neyman-Person detector δ ^NP ( l ), the binary hypothesis defined by formula (2-1)

, assuming that the determined false alarm rate (false alarm rate) P _F is α, the maximum detection capability P _D of the threshold τ _l in the stochastic decision rule in the following formula (2-2) is searched for.

Among them, Λ( z _l ) is defined as Λ ( z _l )= p ₁ ( z _l )/ p ₀ ( z _l ), p ₀ ( z _l ) and p ₁ ( z _l ) can be obtained from formula (2-1) Find out.

Finally, calculate the VD value in the following formula (2-3) according to formula (2-2).

；當δ ^NP(z _l )<1，

。並且，光譜波段數目L介於VD到2VD值之間時，能有較佳的演算分析結果。 Among them, PF is a predefined threshold, when δ ^NP ( z _l )=1,

; when δ ^NP ( z _l )<1,

. Moreover, when the number of spectral bands L is between VD and 2VD, better calculation and analysis results can be obtained.

The "full band analysis method" disclosed in the present invention refers to directly performing calculations on all spectral bands to obtain characteristic bands.

The "Uniform" method disclosed in the present invention is based on the number L of spectral bands determined by the virtual dimension, and uses statistical uniform distribution (Uniform Distribution) to directly analyze all spectral waveforms to be measured (for example, those contained in near-infrared rays) 256 spectral bands) are averaged to obtain the averaged band.

The "Band Selection" (Band Selection, BS) disclosed in the present invention refers to eliminating the section of the spectral waveform to be measured that is highly correlated with the standard spectral waveform through an algorithm, reducing the burden of calculation, and improving the accuracy of the data at the same time. Algorithms that can be used together include Band Prioritization (BP) or/and Band De-Correlation (BD) to select spectral bands suitable for predetermined disease analysis, and then make high-quality Spectral imaging device for subsequent comparative analysis. The band selection correlation calculation is as follows:

Among them, Ω _BS is the set of selected band numbers, J (.) is the optimal band selection objective function, n _BS is the number of selected bands, and Ω ^* _BS is the optimized band subset.

The "band prioritization method" disclosed in the present invention selects a specific band priority criterion (Band Prioritization Criterion, BPC), and gives a priority score according to the importance of information in each band. All bands are then sorted according to their priority scores calculated by the band priority criteria. With these priority scores, bands can be simply added or removed to suit the n _BS value (number of bands selected), and adjustments can be made at any time without recomputing the band selection for the hyperspectral image.

Then, one of the following formulas (5-1) to (5-5) can be used to calculate the variation (Variance), skewness (Skewness), kurtosis (Kurtosis), entropy (Entropy) of different spectral bands respectively ) or Information Divergence (Information Divergence, ID), and as the priority score of the spectral band, where B _j is the jth spectral band, ζ _j is the random variable used to calculate the characteristics of the jth spectral band.

Calculate the variance of the random variable ζ _j :

Among them, σ is the standard deviation (Standard Deviation).

Skewness: Calculate the third-order statistics of the random variable ζ _j :

是計算第j個光譜波段B _j 的三階統計量樣本平均值。 in,

is the third-order statistic sample mean calculated for _the jth spectral band Bj .

Kurtosis: Calculate the fourth-order statistics of the random variable ζ _j

是計算第j個光譜波段B _j 的四階統計量樣本平均值。 in,

is the sample mean of _the fourth-order statistic for computing the jth spectral band Bj .

Entropy: Calculate the randomness of the random variable ζ _j

Among them, p _j =( p _{j 1} ,p _{j 2} , … p _jMN ) ^T is to calculate the probability distribution of the jth spectral band B _j image histogram.

Information Divergence (Information Divergence, ID): Calculate the degree of non-Gaussianization of the band

Among them, p _j =( p _{j 1} ,p _{j 2 ,} p _{j 3} , … ,p _jN ) ^T is to calculate the probability distribution of the jth spectral band B _j image histogram. g _j =( g _{j 1} ,g _{j 2 ,} g _{j 3} , … ,g _jN ) ^T is to calculate the mean and variance of the jth spectral band B _j Gaussian distribution.

The "band de-correlation method" disclosed in the present invention is to convert the band image into a band vector, and the correlation between the vectors can be used to explain the correlation between the two band images. In the embodiment of the present invention The spectral angle matching method (Spectral Angle Mapper, SAM) and spectral information dispersion (Spectral Information Divergence, SID) are used to realize the band de-correlation step. The specific operation steps are as follows:

為要進行去相關性的波段總集合，而b _l 代表lth波段影像B _l 的波段向量。

是波段向量總集合。 Step 1: Order

is the total set of bands to be decorrelated, and b _l represents the band vector of the l th band image B _l .

is the total set of band vectors.

跟n ₁=1 Step 2: Let k = 1 and

with n ₁ =1

Step 3: In the k _th iteration, compute the correlations of all bands b _l in Ω _L − Ω _k .

BD( b _l ,b _k ) (6)

Among them, BD( b _l , b _k ) is used to measure the band correlation between its representative band images B _l and B _k .

And BD can be any calculation method of spectral vector signal, which means that the BD in the formula (6) is set as SID, which can be defined as BD( b _l ,b _k )=SID( b _l ,b _k ), and the SID to calculate the spectral vector correlation between b _l and b _k .

Step 4: If BD( b _l ,b _k ) > ε , then Ω _{k +1} = Ω _k ∪{ b _l } and check that k = L . Otherwise, continue to execute, another Ω _{k +1} = Ω _k and k ← k +1, and then go back to step 3.

Step 5: Let the final result of the iteration Ω _k be Ω _BD represents the final set after decorrelation, where the correlation between all bands in Ω _BD must be greater than ε .

The "abundance scorer" disclosed in the present invention is an algorithm for hyperspectral images, which can mainly calculate the brightness and/or vector of each pixel in the image, including Spectral Matched Filter (SMF) , Adaptive Cosine Estimator (ACE), Constrained Energy Minimization (CEM) or spectral decorrelation energy minimization.

The "spectral decorrelation energy minimization method" disclosed in the present invention is an improvement of the constrained energy minimization method. The main principle is to take a set of L-dimensional hyperspectral image sequences to take the pixel value of the previous point, according to the pixel value of this point A set of L-dimensional vectors can be obtained from the coordinate position, and a set of L-dimensional solutions can be obtained from this set of vectors. The inner product of this set of solutions and the vector of the input coordinates will make the output value equal to 1, and try to suppress other vectors from this set. The value obtained by solving the inner product, so if it is close to the vector of the input coordinates, the output value will be quite close to 1. For grayscale images, it highlights the brightness of the target area and suppresses the non-target area.

Let L be the dimension of Spectral Bands, so the i- th image vector is r _i =( r _{i 1} , r _{i 2} ,..., r _iL ) ^T , and r _ij represents the i -th pixel vector's j spectral bands, { r ₁ , r ₂ ,..., r _N } is the hyperspectral image sequence collection of all pixels in the image, and N is the total number of pixels in the image. Let the spectrum of the target of interest be d, and design an L-dimensional FIR linear filter w =( w ₁ , w ₂ ,..., w _L ) ^T for the target, assuming y _i is the pixel of the i -th spectral image Spectrum ( r _i ) is the output of a designed FIR filter, written as

Calculate the average energy result of all pixels { r ₁ , r ₂ ,..., r _N }:

可以看得出來為一個矩陣自相關的形式，所以，CEM濾波器是解決以下線性約束優化問題：min{w ^T R _L*L w}subject to d ^T w=1 (8) in

It can be seen that it is in the form of a matrix autocorrelation, so the CEM filter is to solve the following linear constraint optimization problem: min{ w ^T R _{L * L} w }subject to d ^T w =1 (8)

Formula (8) its solution:

In order to obtain the severity of diseases, the auto-correlation matrix R _{L × L} in the formula (9) will be slightly improved, and all known standard spectral waveforms will be included in the operation of the auto-correlation matrix, and There is not only a single case, use it as a benchmark for comparison, and perform calculations on abundance images.

It can be seen from the above that the CEM filter can detect δ ^CEM ( r ), when the pixel vector r = d on the hyperspectral image cube, δ ^CEM ( d ) = 1, which satisfies formula (8). Wherein, the value of δ ^CEM ( r ) is obtained through the pixel d to be detected included in the image pixel r . Accordingly, the abundance image was calculated from the CEM filter, and the abundance score of the CEM was normalized by the following formula (10) to calculate the quantitative abundance score representing the severity of the disease for subsequent quantitative analysis.

Furthermore, if the normal (Normal) skin hyperspectral image region is used as the coordinates of interest, this set of solutions can be regarded as a filter so that each pixel vector of the hyperspectral image sequence is sequentially brought into the operation, Complete the detection of normal areas, and suppress the brightness of other abnormal areas, and then achieve target detection with or without disease, and then calculate the severity of the disease (abundance score), and the greater the abundance score It means the more normal, the smaller the value is the abnormal area.

Another example is to take the hyperspectral image region of the most severe level of a predetermined disease (such as diabetes, scleroderma, and rheumatoid arthritis) as the coordinates of interest, and this set of solutions can be regarded as a filter to make the Each pixel vector is sequentially brought into the calculation to highlight the diseased area and suppress the brightness of other normal areas, and then detect the diseased target. At this time, the larger the abundance score represents the abnormal area, the higher the abundance score A small value represents a normal area.

The "electrochemical skin conductance (ESC)" disclosed in the present invention mainly uses reverse iontophoresis (reverse iontophoresis) and chronoamperometry (chronoamperometry) to calculate the concentration of chloride in sweat, thereby detecting The sweat gland function of the subjects perform an accurate assessment. Among them, the detection instrument used in the electrochemical skin conductance detection method has two sets of large-area stainless steel electrodes, which are used to measure the conductance measurement values of the hands and feet of the subject respectively, and the conductance measurement values represent the generated The ratio of the current to a constant DC voltage DC stimulus (≦4V) applied to the electrodes, expressed in microSiemens (μS), and finally, the conductance measurement is transferred to a computer for recording and data management.

In addition, the criteria for judging whether the subject’s feet have motor dysfunction are as follows: a conductance measurement value greater than 60 μS represents no dysfunction; a conductance measurement value of 40 μS to 60 μS represents moderate dysfunction; a conductance measurement value of less than 40 μS represents severe dysfunction. Accordingly, for diabetic patients, the present invention uses the conductance measurement value greater than 60 μS to indicate normal foot nerve function, and the conductance measurement value to be less than or equal to 60 μS to indicate foot nerve function is abnormal, as the basis for grouping.

"The modified Rodnan skin score" (The modified Rodnan skin score, mRSS) disclosed in the present invention divides the human body into 17 common skin parts, such as upper arm, forearm, hand, finger, calf, foot, and abdomen, etc., and then Based on the skin score table, the clinician will actually palpate the skin of these parts to judge the skin scores (Skin scores), or mRSS scores, which will be determined and summed up as the basis for clinical diagnosis.

The "ultrasonic skin thickness detection" disclosed in the present invention is to use an ultrasonic instrument (ie Philips iu22 (Philips Ultrasound, Bothell WA, USA)) to scan the palm side of the bilateral forearm of the selected subject with a 4MHz-10MHz or 15MHz linear probe Ultrasonic images of target lesion areas such as the back of the hand, and then measure the skin thickness in these images.

In one embodiment of the present invention, a method for selecting a hyperspectral feature band is disclosed, which mainly includes the following steps:

Step A: Calculate the number of endmembers of the hyperspectral image of the predetermined disease using the virtual dimension algorithm, and use 1 to 2 times the number of endmembers as the number of predetermined selected bands. The number is defined as n, and n is a natural number other than zero; The predetermined disease may be but not limited to diabetes, scleroderma or rheumatoid arthritis and the like. Wherein, the predetermined disease is scleroderma, and the number of selected bands is preferably 1.5 times the number of endmembers; the predetermined disease is diabetes, and the number of selected wavebands is preferably 1 time or 1.5 times the number of endmembers.

Step B: Obtain most spectral bands from the hyperspectral image, and select n characteristic bands with the same number of selected bands; wherein, the method of selecting characteristic bands can be, but not limited to, all band analysis methods and average methods And the band prioritization method, in simple terms, the whole band analysis method operates on all spectral bands to obtain the characteristic bands.

The average method uses statistical uniform distribution to calculate the average of all spectral bands, and the number of bands is selected to be equal to the number of end members to obtain the characteristic bands.

The band priority sorting method calculates the priority scores of all spectral bands, and sorts the first n spectral bands with higher priority scores as characteristic spectra. Among them, the number of variance, skewness, kurtosis, entropy or information divergence can be selected to calculate the priority score of each spectral band. In addition, when selecting characteristic spectra by band prioritization method, it is better to select the number of bands that is 1.5 times the number of endmembers.

Step C: The system counts the number of repetitions of the same characteristic bands selected by all the band analysis methods, the average method and the band prioritization method, and takes the one whose number of repetitions reaches the predetermined threshold as the better characteristic spectrum.

With the composition of the above steps, the hyperspectral characteristic band selection method disclosed in the present invention can select the characteristic bands of a predetermined disease, so as to achieve the purpose of simplifying the amount of data and streamlining the hyperspectral image analysis steps.

In order to further illustrate the technical characteristics of the present invention and the effects that can be achieved, several embodiments and diagrams will be described in detail as follows.

First of all, in order to clearly distinguish the band selection methods used in the present invention, the names of these methods are redefined, that is, the whole band analysis method is the first band selection method, and the average method is the second band selection method. The band priority sorting method is divided into the third band selection method, the fourth band selection method, and the fifth band selection method based on different band sorting standards, namely, variation, skewness, kurtosis, entropy, and information divergence. method, the sixth band selection method, and the seventh band selection method.

Example 1: Diabetes Detection and Analysis

As shown in Figure 1, the diabetes detection analysis mainly includes the following steps.

Step 101: Recruiting subjects

Taichung Veterans General Hospital in Taiwan recruited 176 diabetic patients and 20 healthy subjects.

Step 102: Detection of electrochemical skin conductance

The subjects were tested for electrochemical skin conductance and grouped according to the test results, as shown in Table 1 below.

Step 103: Obtain hyperspectral image information

The skin hyperspectral images of the feet of each subject were taken, and image pre-processing was performed separately. For example, the image pre-processing method in this example is to use two images of different bands to perform band ratio (Band ratio), and then use the morphology of traditional spatial image processing, combined with Otsu thresholding method (Otsu thresholding ), the foot area in the hyperspectral image of each subject can be extracted and made into a mask. Among them, as shown in Figure 2 to Figure 4, they represent the general image, hyperspectral image, and subsequent extraction of the mask image of the subject's foot, respectively, and the mask image can be used to automatically map the foot All pixels in the inner region are taken out for hyperspectral image analysis.

Then, the hyperspectral waveform difference analysis method is used to calculate the corresponding spectral waveforms respectively.

Step 104: Determine the number of endmembers

The number of endmembers (VD value) is calculated as 14 by the virtual dimension algorithm, and the number of endmembers (VD value) is used as the predetermined number of bands. In particular, when selecting a band, it is not always possible to find the number of bands that exactly matches the number of endmembers (VD value), so the number of bands can be between VD and 2VD.

Step 105: Band Selection

As shown in Table 2, the band selection method adopted, the number of bands and the selected characteristic bands are listed in series. In FIG. 5A to FIG. 5D , there are distribution diagrams of bands selected by different band selection methods, and the spectral wavelength positions of the selected bands are marked.

Among them, when band analysis is carried out with the third band selection method, the fourth band selection method, the fifth band selection method and the sixth band selection method, the number of bands is based on the value close to 1.5VD (at this time, _PF = 10 ^-4 ), and the second band selection rule is to analyze at 1VD.

From the above analysis results in Table 1, it can be seen that the number of characteristic bands selected by different band selection methods is at least 14. However, if all 14 spectral bands are included in the reference basis for the construction of an active hyperspectral imaging device A, it is impossible to This device shrinks again, is difficult to reach the main purpose of the present invention.

Therefore, under the consideration of the hardware production of the photographable range, further count the number of occurrences of the characteristic bands selected by different band selection methods, and find out the characteristic bands selected by each method that appear the most times. The results are shown in the table 3.

According to the content of Table 3, among the characteristic bands selected by different methods, the bands that repeat three times are 1125.4nm, 1155.7nm, 1373.4nm, 1386.7nm, 1393.4nm, 1648.6nm, 1655.1nm, 7 in total. However, considering factors such as current short-wave infrared LED products, prices, and functions, 1100nm, 1150nm, 1200nm, 1300nm, 1450nm, 1550nm, and 1650nm were finally selected as the specific light source 20 of the active hyperspectral imaging device A.

Step 106: Verify comparison

In order to verify the calculation results of different band selection methods, a Hyperspectral Imaging Soft Abundance Scorer (HISAS) constructed by the spectral decorrelation energy minimization method was used to match the known disease spectrum band, and the hyperspectral image information of the most severe level of diabetes as a calculation basis, and calculate each pixel vector in each spectral band, so as to calculate the first band selection method, the second band selection method, the third band selection method, Abundance Fractions/Abundance Score or Abundance Fraction Map of the fourth band selection method, the fifth band selection method, and the sixth band selection method, and the value of the abundance score can be used to judge whether There is no evidence of diabetes.

Then, use the receiver operating characteristic curve (ROC curve) to compare the abundance scores calculated by these methods, as shown in Figure 6, which shows the analysis results of these methods to determine whether there is diabetic foot neuropathy The ROC curve diagram.

Then, the area under the ROC curve (Area Under Curv; AUC) of these methods was calculated separately, and the AUC value was used as a reference for the diagnostic ability. Then, the AUC value of each method was compared with the electrochemical skin conductance test. The results are compared and verified, that is, the foot nerve function is normal if the conductance measurement value is greater than 60 μS, and the foot nerve function is abnormal when the conductance measurement value is less than or equal to 60 μS, and the correlation between each method and the conductance test results The results of the sex analysis are presented in Table 4.

It can be seen from the above table that the calculation results of the area under the ROC curve of the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, and the sixth band selection method are respectively They are 0.857, 0.866, 0.870, 0.871, 0.873, 0.876, among which, the diagnostic ability of the sixth band selection method is the highest, that is, the AUC value is 0.876. Moreover, except for the correlation between the second band selection method and the conductance test result which is 0.579, the correlations between the other methods and the conductance test result are all above 0.58.

Example 2: Detection and analysis of scleroderma

As shown in Figure 7, the main implementation steps of the scleroderma detection analysis include:

Step 201: Recruit subjects

Taichung Veterans General Hospital in Taiwan recruited 30 patients with scleroderma (average age 49.93±17.01, 10 males and 20 females) and 24 healthy individuals (average age 37.01±10.97, 12 males and 12 females). For scleroderma patients over 20 years old, and approved by the Human Experiment Review Board (IRB), the human experiment research plan license numbers are CE16035B and CE16201A. safety and privacy of the user.

Step 202: Calculate skin score and detect skin thickness

Clinicians used the skin score sheet to diagnose the subjects' skin assessment scores, and measured the subjects' skin thickness with ultrasonic equipment for subsequent correlation analysis, and the data can be found in Yi-Ming Chen Equal to "Hyperspectral" published in the journal Rheumatology in 2020 Imaging for skin assessment insystemic sclerosis: a pilot study" Table 1 and Figure 1 in the paper.

Wherein, this analysis and measurement method is a traditional technology, so its detailed information is not repeated.

Step 203: Obtain hyperspectral image information

The hyperspectral imager was used to capture the skin information of six different target areas on the subject's hands (ie, 3 for the left hand and 3 for the right hand). The shooting time was about 10 minutes, and hyperspectral images were obtained.

Step 204: Determine the number of endmembers

The number of endmembers (VD value) is calculated to be 20 for the hyperspectral images of all subjects using the virtual dimension algorithm, and the number of endmembers (VD value) is used as the predetermined number of bands. In particular, when selecting a band, it is not always possible to find the number of bands that exactly matches the number of endmembers (VD value), so the number of bands can be between VD and 2VD.

Step 205: Band selection

As shown in Table 4, the number of bands determined by all methods and the characteristic bands selected by them are listed in series. In FIG. 8A to FIG. 8F , there are distribution diagrams of bands selected by different band selection methods, and the spectral wavelength positions of the selected bands are marked.

Wherein, when the average band and the hyperspectral characteristic band selection method are used for band selection, the number of bands is based on a value close to 1.5VD ( _PF =10 ^-4 at this time).

It can be found from the above table 5 that among the characteristic bands selected by the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method, and the seventh band selection method, they are sorted in the first The band numbers (wavelengths) of the spectral bands are 68 (1085nm), 182 (1467nm), 175 (1443nm), 195 (1510nm), and 86 (1146nm). Moreover, the distribution of spectral bands picked out by different methods is very average, indicating that these methods are equally capable of judging the presence or absence of scleroderma.

Furthermore, in Table 5, the characteristic bands selected by the fourth band selection method, the fifth band selection method, and the sixth band selection method have 7 identical spectral bands, that is, the band numbers (wavelengths) are 31 ( 960nm), 82(1132nm), 84(1139nm), 87(1149nm), 89(1156nm), 91(1162nm), and 111(1230nm), although the sorting results of these spectral bands are different.

In addition, for biological tissues, water (water) has a wavelength of 1450nm, and lipids (lipid) and collagen (collagen) have a wavelength of 1100nm. There are obvious differences in the vicinity of these two spectral bands, and the methods selected Among the characteristic bands that come out, there are spectral bands close to 1100nm and 1450nm. Therefore, 1100nm and 1450nm can be used as the basis for scleroderma detection.

Step 206: Correlation coefficient analysis and comparison

Then, the hyperspectral image soft abundance scorer designed by the improved restricted energy minimization method is used to calculate the soft abundance score of the disease degree for these characteristic bands, and finally compare with the clinical data respectively. According to (such as skin thickness) and the results of evaluating the severity of the disease by spectral difference analysis in the aforementioned literature, a statistical comparison was performed.

As shown in Figure 9A to Figure 9G, the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, and the sixth band selection method adopted in the present invention The box plot of the statistical analysis results of the Mann-Whitney U test (Mann-Whitney U test) for the left and right skin softness abundance scores of healthy subjects and scleroderma patients calculated by methods such as the Seventh Band Selection Method , among which, the average abundance score of the hand skin of healthy people is close to 1, while the average abundance score of the hand skin of patients with scleroderma is around 0.5.

However, as shown in FIG. 9H , it is a box plot of the statistical analysis of the conclusions of the aforementioned literatures using the Mann-Whitney U test, in which the average value of healthy people is about 0.019, while the average value of scleroderma is about 0.22.

Because both the present invention and the aforementioned documents use the hyperspectral waveform of healthy skin as the analysis benchmark, but the results are very different, the main reason is that the hyperspectral image soft abundance scorer proposed by the present invention can be regarded as a filter, It will suppress the signals that are different from the comparison reference spectral waveform, and highlight the desired target waveform signal.

On the other hand, the spectral difference analysis method adopted in the above-mentioned literature only simply compares the difference between two spectral waveforms. Although the statistical results of the two groups of healthy people and scleroderma patients are significantly different, it still results in the analysis of the average value of scleroderma skin. The results were similar to those of healthy subjects.

Next, the results of soft abundance scores of all methods were statistically compared, and the results were all significantly different, as shown in Table 6.

Mann-Whitney U test, interquartile range Median (IQR). Among them, *p<0.05, **p<0.01, and * mark is significant, ** mark is very significant.

Furthermore, it has been preliminarily proved in the aforementioned literature that the hyperspectral image analysis technology is superior to the current clinical diagnosis method (ie, skin thickness detection) in assessing the judgment of the disease in patients with scleroderma. Therefore, with the first band selection method adopted in the present invention, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method, the seventh band selection method, etc. The method was compared with the literature mentioned above.

As shown in Table 7, it is the comparison result of Spearman's rank correlation coefficient (Spearman's rank correlation coefficient) between these methods. Wherein, because the present invention is the hyperspectral waveform of the healthy person as a comparison reference, and the smaller the calculation result value of the spectral difference analysis method, the closer to the hyperspectral waveform of the healthy person, so the spectral difference analysis method and different band selection Methods All analysis results are negatively correlated, that is, the Spearman rank correlation coefficients are above -0.631.

The first band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method, and the seventh band selection method all have a Spearman rank correlation coefficient as high as 0.965 or more.

Spearman's rank correlation coefficient, where, * p <0.05, ** p <0.01, and * mark is significant, ** mark is very significant.

As can be seen from the foregoing, the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method, and the seventh band selection method adopted in the present invention etc. methods are respectively superior to the spectral difference analysis method adopted in the foregoing literature, and its main possible reason is that the spectral difference analysis method only simply considers the difference of spectral vectors, and the present invention utilizes the improved When using the formula-limited energy minimization method for sub-pixel detection, the autocorrelation matrix in the calculation formula takes into account the relationship between the sample data, so the analysis result is better than that of the spectral difference analysis method.

Furthermore, it can be found from the contents of Table 7 that the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, and the sixth band selection method adopted in the present invention , the seventh band selection method and other methods have a very high correlation, which means that the selected characteristic bands are all representative, so there will be no difference in analysis.

In addition, we can also find that the correlation coefficient among the second band selection method, the third band selection method, and the fourth band selection method in Table 7 is as high as 0.997 or more. The possible reason is that the essence of these three methods are all In describing the relationship between the high-order statistics (high-order) of the data, there is such a high correlation coefficient.

Step 207: ROC curve comparison analysis

In order to understand the first band selection method, the second band selection method, the third band selection method, the fourth band selection method, the fifth band selection method, the sixth band selection method, the seventh band selection method and other methods adopted in the present invention The ability to discriminate scleroderma was evaluated by using the operator acceptance curve (Receiver operating characteristic curve, ROC curve) to evaluate the abundance scores calculated by each method, as shown in FIG. 10A to FIG. 10G . Then, it was compared with the skin score and skin thickness of the hands of scleroderma evaluated by clinicians, and the analysis results of the aforementioned literature, as shown in Table 8.

In addition, from Fig. 10A to Fig. 10H and Table 8, it can be found that the area under the curve or the accuracy (Accuracy) of each of the methods adopted in the present invention is far better than the clinical skin thickness measurement method, and is also better than the aforementioned The spectral difference analysis method in the literature is better than or equal to the skin evaluation score results of clinicians.

Among them, the spectral difference analysis method is not consistent with the analysis results of the above-mentioned literature, which may be caused by the different number of samples and the characteristics of the spectral difference analysis method itself. In detail, the sample ratio of scleroderma patients to healthy subjects in the aforementioned literature is 31:19, while the sample ratio of scleroderma patients to healthy subjects in the present invention is 30:24. Moreover, the spectral difference analysis method only compares the difference between two spectra, so if there is some noise interference (such as skin dandruff, moles, or broken capillaries, etc.) Due to the difference of the spectrum, the results of the analysis based on the spectral difference analysis method, when classifying whether there is scleroderma, the sensitivity (Sensitivity) and specificity (Specificity) decrease, which is different from the analysis results of the aforementioned literature.

Therefore, the present invention further proposes a hyperspectral image soft abundance scorer to improve the problem that the traditional spectral difference analysis method cannot consider the correlation of spectral waveforms between all skins, and uses the constrained energy minimization method (CEM) to perform the soft scorer During the analysis, the noise interference problem will be suppressed, so that the noise interference can be reduced, and the deviation in the statistical correlation analysis will not be caused.

In summary, the present invention uses a hyperspectral image softness abundance scorer to analyze the hyperspectral image of scleroderma hand skin, and the calculated softness abundance score can be used as a judgment of the degree of scleroderma. Furthermore, the present invention discusses the bands of hyperspectral images of scleroderma hand skin based on high-order statistics-based and infinite-order statistics-based band sorting methods From the analysis results, it can be seen that most of the characteristic bands selected by different band selection methods are similar, and it is quite good to use the abundance scores calculated after finding out these bands to judge the severity of the disease. It is a reference opinion that can help clinicians to carry out diagnostic assistance.

In addition, these characteristic wavebands selected by the present invention can be used to improve traditional hyperspectral imagers.

As shown in FIGS. 11 to 13 , another embodiment of the present invention discloses an active hyperspectral imaging device A, which mainly includes a base 10 , a plurality of light sources 20 and a sensing part 30 .

The base 10 has a housing 11, a cover 12 and a base 13, wherein the housing 11 has a housing body 111, and an assembly port 112 opened in the housing body 111, so that the housing body 111 The internal space 113 communicates with the outside world.

The cover body 12 has a plate-shaped cover body 121, which is connected with the shell body 111 corresponding to the assembly port 112, and a first hole 122 of an appropriate aperture runs through the central position of the cover body 121, and A recess 123 with a suitable depth is recessed on one side of the cover body 121 and communicates with the first hole 122 .

The seat body 13 is provided with an annular seat body 131 and a second hole 132 of a suitable diameter, which is axially penetrated on the seat body 131 along the center of the seat body 131, and the seat body 131 is The second hole 132 is coaxial with the first hole 122 to be removably embedded in the recess 123 .

電機株式会社(USHIO,Inc.,Tokyo,Japan)或其台灣子公司優志旺股份有限公司(USHIO TAIWAN,INC.)所生產之短 波紅外線(Short Wave InfraRed，SWIR)LED燈，其中，短波紅外線LED燈的產品種類繁多，大致上可分為SMBB、EDC、SMT及SMC家族等系列，而基於硬體結構設計、價格、LED晶片功能及LED光源發射特性等因素考量下，本實施例係採用EDC系列的LED產品，其單顆LED產品的面積為3.5mm x 3.5mm。 Each of the 20 light sources is selected from Japan

Short-wave infrared (Short Wave InfraRed, SWIR) LED lamps produced by USHIO, Inc., Tokyo, Japan or its Taiwan subsidiary, USHIO TAIWAN, INC., among them, short-wave infrared LED There are many types of lamp products, which can be roughly divided into SMBB, EDC, SMT and SMC series. Based on factors such as hardware structure design, price, LED chip function and LED light source emission characteristics, this embodiment adopts EDC series of LED products, the area of a single LED product is 3.5mm x 3.5mm.

In particular, the wave bands emitted by each of the light sources 20 are respectively configured according to the characteristic wave bands or better characteristic wave bands calculated by the aforementioned hyperspectral feature band selection method. For diabetes, according to the conclusion of Example 1, these The spectral wavebands of the light source 20 can be selected as 1100nm, 1150nm, 1200nm, 1300nm, 1450nm, 1550nm, and 1650nm, a total of 7 kinds of wavebands.

Furthermore, the number of these light sources 20 is 56, and each of the 8 light sources adopts the same spectral band, and are respectively arranged on the seat body 131 in a ring shape with the second hole 132 as the center, as shown in FIG. 13 Show.

Furthermore, as the power of the LED increases and the heat dissipation efficiency increases, the light decay will be reduced and the service life will be prolonged. Therefore, the seat body 131 is made of porous ceramic molding technology, which can provide more air contact area , in addition to good heat conduction and heat dissipation performance, on the basis of this advantage, the flexibility of circuit design or heat dissipation structure design of the active hyperspectral imaging device A is improved.

In the present invention, the light sources 20 are arranged on the seat body 131 to form a set of detection light source components for diabetes, and are suitable for being embedded into the recess 123 of the cover body 12 as a fixing means. In other embodiments, the detection light source assembly can also be positioned on the cover 12 by means of screw lock or magnetic attraction.

In addition, the active hyperspectral imaging device A of the present invention is not only capable of detecting diabetes, but also applicable to other diseases, such as the detection of diseases in patients with scleroderma, and only needs to replace the detection light source assembly, which can be quickly and conveniently to detect scleroderma. Wherein, for these light sources 20 included in the detection light source assembly to be replaced, the selected spectral bands can be calculated for scleroderma with the aforementioned hyperspectral characteristic band selection method, and according to the conclusion of Example 2, these light sources The spectrum band of 20 can be selected as 1100nm and 1450nm, a total of 2 kinds of bands.

It can be seen that the detection light source assembly can be replaced quickly and easily, and can correspond to different predetermined diseases, which is quite convenient in actual operation.

The sensing part 30 has a main body 31 , a cylindrical image capturing lens 32 and a processing unit 33 , wherein the main body 31 is disposed in the inner space 113 of the casing 11 . The image capture lens 32 can be but not limited to a video camera, a camera, a device including charge-coupled device (CCD) or complementary metal oxide semiconductor (CMOS), and is arranged on the body 31, and is used to pass through the first A hole 122 and the second hole 132 make one end of the cylinder shaft of the lens exposed outside the casing 11, and are used to receive the reflected light of an external object irradiated by the light sources 20 to obtain a hyperspectral image, and the other end of the column shaft of the lens is located in the inner space 113 of the casing 11 . Wherein, the target refers to any one of the subject's hands, arms, face, neck, feet or other parts.

The processing unit 33 is set on the main body 31 and is electrically connected with the image capturing lens 32 to receive the hyperspectral image and perform analysis and calculation, and obtain the judgment result of the target object in a predetermined disease, or other Diagnostic information.

According to the above structural description, the active hyperspectral imaging device A of the present invention is equipped with a light source 20 corresponding to the characteristic band of the predetermined disease. Compared with the traditional hyperspectral imaging device, it is no longer necessary to reserve filters and spectroscopes. and other components, and save the configuration of tubes and wires of the ring-type halogen lamp, so that its volume can be reduced to within 18cm x18cm x 20cm. According to this, it can achieve the purpose of lightweight and hand-held design, in order to scientifically evaluate Early detection of diabetic foot disease provides clinicians with more objective diagnostic information for diabetic foot patients.

As shown in FIG. 14 , it is another specific embodiment of the present invention. The main system is to set the active hyperspectral imaging device A on a frame 40 so that it has a working distance of 50 cm from the target object. And the distance between the light sources 20 and the object is 45 cm, and has the largest shooting range, ie, 30 cm x 30 cm.

Next, carry out the intensity and uniformity test of the light source 20, first take pictures of the white paper, and use the center of the image as the benchmark for measuring the intensity, that is, the grayscale value of the center of the image is normalized (normalized) to 100%, and the closer to the center The grayscale value is the highest, and the farther away from the center, the lower the grayscale value, so as to convert the intensity attenuation degree of the light source 20 .

Finally, the above-mentioned 7 wave bands (ie 1100nm, 1150nm, 1200nm, 1300nm, 1450nm, 1550nm, and 1650nm) were tested respectively. The intensity and uniformity of each wave band can reach more than 80% within the range of 25 cm x 25 cm. Within the range of 30 cm x 30 cm, the strength and uniformity can reach more than 60%, as shown in Figure 15A to Figure 15G.

Claims (11)

A method for selecting a hyperspectral characteristic band, comprising the following steps: calculating the number of endmembers of a hyperspectral image of a predetermined disease with a virtual dimension algorithm, and using 1 to 2 times the number of endmembers as the predetermined number of selected bands, The number is defined as n, and n is a natural number that is not zero; then, from the majority of spectral bands contained in the hyperspectral image, select n bands that are similar to the aforementioned selected bands by using the full band analysis method, the average method, and the band prioritization method. The number of characteristic bands with the same number; then count the number of repetitions of the same characteristic bands selected by the analysis method of all bands, the average method and the band prioritization method, and take the one with the same number of repetitions as the predetermined threshold value as the better characteristic spectrum. The method for selecting a hyperspectral feature band according to Claim 1, wherein the predetermined disease is scleroderma, and the number of bands to be selected is preferably 1.5 times the number of endmembers. The method for selecting hyperspectral characteristic bands as described in Claim 1, wherein the priority scores of all spectral bands are calculated using the band prioritization method, and the first n spectral bands with higher priority scores are used as characteristic spectra. The hyperspectral feature band selection method as described in Claim 3, wherein the band prioritization method selects the variance, skewness, kurtosis, entropy or information divergence to calculate the priority score of each spectral band. The method for selecting bands of hyperspectral features as described in Claim 3, wherein the predetermined disease is diabetes, and when performing calculation and analysis with the average method, the number of selected bands is preferably 1 times the number of endmembers. The method for selecting bands of hyperspectral features as described in Claim 3, wherein the predetermined disease is diabetes, and when the band prioritization method is used for calculation and analysis, the number of selected bands is preferably 1.5 times the number of endmembers. An active hyperspectral imaging device, comprising: a base; a plurality of light sources are respectively arranged on the base, and the wavelength bands emitted by each of the light sources are respectively according to the characteristic bands described in any one of claims 1 to 6, or the preferred characteristic wavelength band; and a sensing part, which is separated from the light sources and set on the base, is used to receive the reflected light of an external target irradiated by the light sources to obtain A hyperspectral image. The active hyperspectral imaging device as claimed in item 7, wherein the light sources are arranged on the base in a circular arrangement around the position of the sensing part. The active hyperspectral imaging device as described in claim 8, wherein the base includes: a housing, which has a housing body, and an assembly port opened in the housing body, so that the internal space of the housing body communicates with the outside world ; A cover body, which has a plate-shaped cover body, which is connected to the shell body corresponding to the assembly port, a first hole with an appropriate aperture runs through the central position of the cover body, and a suitable depth The recess is recessed on one side of the cover and communicates with the first hole; and the seat has a ring-shaped seat and a second hole with an appropriate hole diameter along the seat. The center of the body is axially installed on the seat body, and the seat body is used to be embedded in the recess, and the second hole is coaxial with the first hole; wherein, the light sources are respectively connected to the first hole. The two holes are the center of the circle, and are arranged in a ring on the body of the seat. The active hyperspectral imaging device as described in Claim 9, wherein the sensing unit includes: a body, which is arranged in the inner space of the housing; a cylindrical image capturing lens, which is arranged on the body, and used passing through the first hole and the second hole, so that one end of the column shaft of the lens is exposed outside the casing, and the other end of the column shaft of the lens is located in the inner space of the casing; and A processing unit is arranged on the main body and electrically connected with the image capturing lens. The active hyperspectral imaging device as claimed in item 9, wherein the base is made of ceramic material.

Images