TWI537762B - Application of hyperfamily imaging to identification of cancerous lesions - Google Patents

Description

Method for identifying cancer lesions by using hyperspectral image

This creation is about a method for identifying cancerous lesions, especially the application of hyperspectral imaging and principal component analysis to identify cancerous lesions.

With the maturity of Hyperspectral Image Technique imaging technology, hyperspectral imaging technology began to be applied to medical detection. For example, initial oral cancer detection, enterovirus oral lesion detection, or rectal mucosal detection. There are different types of hyperspectral imaging technology depending on the instrument and equipment. Among them, an existing hyperspectral imaging system uses a single-point spectrum analyzer with a two-dimensional scanning system. Although this hyperspectral imaging system can obtain the best analog spectrum and spatial resolution, it takes a long time to read. data. Another hyperspectral imaging system uses a digital camera with a liquid crystal-coordinated filter and a microscope, which can be applied to the detection of bone marrow cells. Although this method can classify the components in bone marrow cells, it is affected by the liquid crystal-coordinated filter. The impact of control limits the speed at which spectral data is read. In addition, another hyperspectral imaging system uses a hyperspectral camera for spectrum and image analysis. It has been applied to beauty and skin detection. Although it has a very high imaging resolution, it needs to process huge data and high cost.

Early detection and diagnosis of endoscopy is the key to reducing mortality. However, the endoscopic technique used in medical treatment to detect early esophageal cancer is white endoscopy. White light endoscopy detection has three characteristic changes according to the early esophageal cancer mucosa: (1) mucosal color changes, there are two forms of red and white. The red color changes to a red area with a clear boundary, the mucous membrane is slightly rough and turbid, and a few are large red areas with unclear borders; white changes are mucous white spots, which are scattered, clear boundaries, varying in size, rough, dull, slightly (2) Mucosal thickening and changes in vascular structure: normal esophageal mucosa epithelium is translucent, submucosal vascular network is clearly visible, and mucosal epithelial carcinogenesis, vascular network is not transparent; (3) mucosal morphology Changes, including erosion, plaque, roughness, and nodules, are mixed, and all three of these changes will cause the esophageal mucosa to lose its normal structure and luster. However, the conventional esophageal mucosa cannot clearly observe the esophageal mucosa. The fine structure must be diagnosed by means of biopsy and staining techniques. The method of staining endoscopic surgery refers to the comparison of the color of the lesion with the normal mucosa through oral, injection, and direct spraying of the dye, thereby facilitating the identification of cancer and the target biopsy, and improving the diagnosis of early esophageal cancer detection. Dyeing endoscopy has the following forms, such as iodine staining, because the normal esophageal mucosa belongs to Squamous Cell Epithelium, which contains many glycogen (Glycogen), because hepatic sugar has strong affinity for iodine solution. It will be stained brown; on the contrary, once the mucosa is lesioned, the glycogen-rich cells will be reduced or disappeared, so it is not easily stained with iodine solution, so according to this principle, when performing endoscopy, Iodine solution can be sprayed on the surface of the esophagus. If it is found that the area where iodine cannot be stained, that is, it is highly suspected to be the possibility of early esophageal cancer, it is the most common stain endoscopy. Toluidine blue staining method, which is an eosinophilic dyeing solution, has affinity for cancer cells and cells before cancer, and can be used for detecting cancerous lesions and cancerous changes. Methylene blue staining is an absorbent dye. Normal esophageal squamous cells do not absorb methylene blue without staining, but can be absorbed by intestinal cells and columnar cells and stained blue. Therefore, they are commonly used in esophageal glands. The detection of cancer, but in the medical literature shows that methylene blue may cause DNA damage, and time-consuming depends on the operator's experience. Acetic acid staining method, which can make the cytoplasmic protein in the cell produce reversible changes. The commonly used concentration is 1.5~3%. After spraying for 2-3 minutes, the esophageal squamous epithelium still maintains white color, while the columnar epithelium turns red. Identify residual columnar cells. Due to the different distribution of dye concentration, the limitation of the spraying method, the limitation of the dyeing agent itself, etc., the staining depth, the inaccurate positioning or the omission of the lesion may be caused by the staining, and the single staining method has certain limitations. Therefore, it can be improved by double staining.

However, in addition to the above-mentioned chemical dyeing endoscopy, there are electronic dye imaging techniques by optical changes, such as Narrow Band Imaging (NBI), Intelligent Electron Molecular Spectroscopic Endoscopy System (Fujinon Intelligent Chromoendoscopy). , FICE), both methods are based on the choice of a certain wavelength spectrum, but NBI is to use the filter to narrow the spectrum bandwidth, while FICE is to decompose the traditional white light image into many single wavelength spectral images, and then from Capture and synthesize images of the appropriate wavelength.

In view of the difficulty of endoscopic detection, there are two medically sensitive and convenient clinical methods to identify benign or malignant lesions, namely Multiphoton Microscopy (MPM), and with the biological sciences. Progressively, surface-enhanced Raman Scattering (SERS) using nanoparticle polymers is also applied to the detection of cancer cells. Two-Photon Excited Fluorescence (TPEF) and Second Harmonic Generation (SHG) signals in multiphoton microscopy (MPM) technology, from background signals to autofluorescence Differences were made to identify esophageal cancer lesions. As shown in Figure 8, it can be seen from the signal intensity ratio of SHG and TPEF that the cancerous tissue is significantly lower than normal tissue, which clearly indicates the collagen content, distribution, and morphological structure, because cancer cells occupy submucosal tissue. Changes occur, so the signal intensity ratio can be used as a quantitative indicator to distinguish between normal, precancerous lesions and esophageal cancer tissue. On the other hand, Raman spectroscopy provides information on the molecular composition and specific fingerprint forms of biological tissue structures, and is a viable optical technique for diagnosing and evaluating cancer. However, there are two main defects in the general Raman spectroscopy technique, that is, the background fluorescence interference is too strong, and the efficiency of Raman scattering is too low, which makes the practical clinical application difficult, so surface enhanced Raman spectroscopy (SERS) can solve these problems. The weak Raman signal can be increased to the 14th power of 10, and the background fluorescent signal is also greatly reduced by the molecular absorption on the surface of the nanoparticle metal, as shown in Figure 9, in the absence of nano silver particles. When the sample of esophageal cancer tissue to be tested is dropped, the Raman signal is quite weak. As shown in the Raman spectrum of Fig. 9, the nano silver particles with an average diameter of about 25 nm are dropped into the esophageal cancer tissue for two minutes. Its Raman signal strength is greatly improved, as shown in the surface-enhanced Raman spectrum of Figure 9. Finally, through the results shown in FIG. 10, the surface-enhanced Raman spectroscopy signals of the normal tissue waveform and the cancerous tissue waveform are compared, and the waveform difference spectrum can be seen.

However, early lesions are not apparent under white light and are easily overlooked to delay treatment. With the advancement of endoscopic techniques in recent years, the use of endoscopes combined with some chemical and optical principles can highlight small esophageal cancer lesions and improve the diagnosis rate of early cancer, such as iodine staining endoscopy and narrow frequency. Narrow-band Imaging (NBI). However, in the process of iodine-stained endoscopy, the dyeing agent is sprayed to achieve the development effect, and the uneven distribution of the dyeing agent may cause difficulty in discrimination. In addition, the dye may cause irritations such as stinging or burning in the chest of the patient. . NBI has some subjectivity for image analysis, and it is affected by some factors. For example, if there is bleeding or inflammatory lesions, it will result in blurred vision, poor image resolution, and significant organ dysfunction in patients. For example, the heart, lung disease, and unstable conditions must also be evaluated in detail for their organ function and whether the balance is suitable for examination. At present, in the detection of esophageal cancer, the NBI technology is currently the mainstream, because the dyeing effect can be achieved without spraying the dye, and the operation is quite easy, and the white light can be switched to the NBI mode at the push of a button. Repeated observation, however, the basis of the test is mainly to observe changes in the superficial blood vessels such as the intra-epithelial Papillary Capillary Loop (IPCL), but it relies heavily on the subjective judgment of the clinician.

The working principle of the aforementioned NBI is to use a special filter to filter out two narrow-band blue and green light (415 nm and 540 nm) by Xenon. However, due to the early development of esophageal cancer, it is often accompanied by The surface of the mucosa is hyperplasia, and the depth of visible light wavelength penetration is deeper according to the wavelength. The deeper the penetration depth, the blue light is discarded, and only the blue-and-green light is considered. Then the photosensitive-coupled device (CCD) receives To the narrow-band blue and green light incident on the mucous membrane, convert it into a digital signal and assign the digital signal to the R, G, B channels according to the color redistribution of human color vision, ie blue light (415 nm) is assigned to the B and G channels; green light (540 nm) is assigned to the R channel, and finally, through the color conversion process, the endoscope image originally having only the blue and green components can display a color image on the screen. As shown in Annex 6, it can be seen that the blood vessels in the surface layer will appear brown, while the blood vessels in the deeper layers will appear green. This way, the fine blood vessels on the surface of the mucosa will be sharper and more contrasting. If you use the magnifying endoscope together, you can magnify the lesion by more than 80 times, so that the clinician can clearly observe the changes in the texture, microvascular arrangement and thickness of the lesion surface. Therefore, by observing the difference of these brown lesions, It can effectively improve the diagnosis rate of early esophageal cancer.

Principle Component Analysia (Principal Component Ananaysis) is a commonly used method for multivariate statistics. Since 1960, it has been applied to people in color technology. The concept of principal component analysis is to find the ratio in a multivariate data set. The original variables are small, and the subspace of the data changes can be retained, and the original data is projected to this subspace analysis. The main purpose of its principal component analysis can be divided into two: the first is to define the main axis direction of a large amount of spectrum information, and the second is to simplify the data of the information, mainly after reorganizing the original data, and calculating the correlation and Variables that are independent of each other, and then the main components are obtained by analysis, and finally the variability of most of the data in the original data can be explained. The example in Figure 11 explains the geometric meaning of the principal component analysis. In Figure 11, there are two groups of data points with uneven distribution. From the X1 and Y1 coordinate systems, the two points are separated from the X1 or Y1 coordinates. Group information is not easy. Because the data from these two directions, there is no obvious boundary area. After the principal component analysis operation, a spindle direction can be obtained as shown by X2. Therefore, after projecting the data to X2, all the data have a new coordinate value X2. Obviously, the data is obviously different on X2, which is the characteristic of principal component analysis.

Therefore, if a new optical detection method can be provided by using the principle of hyperspectral imaging technology and principal component analysis, the IPCL pattern change of the esophageal cancer lesion in the NBI endoscope image, and the esophagus in the white light and iodine stained endoscope image are normal, The spectral characteristics of precancerous lesions and cancerous lesions, comparing the results of their main component score maps and their spectral characteristics, can help clinicians quickly identify early cancerous lesions in the esophagus.

The purpose of this creation is to provide a method for identifying cancer lesions using hyperspectral imagery to quickly assess the likelihood of a patient's staging in each cancer.

According to the above object, the present invention provides a method for identifying a cancer lesion by using a hyperspectral image, comprising the following steps: obtaining a plurality of first pathological images from an endoscope, the plurality of first pathological images being a plurality of cancer lesion images; The first pathological images are imported into an image processing module to obtain a plurality of first simulated spectra of the first pathological images, and a principal component score map is drawn according to the first simulated spectra; Simulating a spectrum, defining a plurality of triangle ranges in the principal component score map; determining whether a principal component score of a second analog spectrum of a second pathology image falls within one of the triangle ranges; when the second analog spectrum The main component score falls within one of the triangles, and the second pathological image is confirmed to belong to some of the cancer lesion images.

Through the application of the hyperspectral image recognition cancer pathology method, the cancer lesion image can be digitized, and the principal component analysis can be used to effectively and quickly improve the diagnosis efficiency of the doctor and help the patient to perform early treatment.

FIG. 1 is a flow chart of an embodiment of a method for identifying a cancer lesion using a hyperspectral image. As shown in FIG. 1, in step S101, a plurality of first pathological images are obtained from an endoscope, and the first pathological images are cancer pathological images. The hyperspectral imaging system constructs an endoscope mainframe and a high-resolution spectrometer. The hyperspectral imaging system captures X-Rite (Mini Color Checkers) image information. In order to determine the cancer, to obtain the spectrum of each pixel in each first pathological image, it is necessary to first find the relationship matrix between the spectrometer and the endoscope. The spectrum of the 24 color patches needs to be measured by an optical spectrometer in an environment of an endoscope, and the spectrum is set in the visible light band (380 nm to 780 nm). For the convenience of analysis, these spectra are organized into a matrix of 401*24, each column of the matrix is the intensity value corresponding to the wavelength, and each row represents the number of color patches.

These 24 color patches are simultaneously captured in an endoscope environment. These 24-color block output formats are sRGB (data for JPEG images). By computer calculation, you can get the red (Red, R), green (Green, G), blue (Blue, B) values (0 ~ 255) in each color block, and convert to a smaller scale ( 0~1) R _srgb , G _srgb and B _srgb . With the following formula, these RGB values will be converted into tristimulus values X, Y, Z under the International Commission on Illumination (CIE) specifications, as follows:

<TABLE border="1" borderColor="#000000" width="_0001"><TBODY><tr><td width="183" height="5"></td></tr><tr>< Td></td><td><img wi="135" he="77" file="02_image001.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(1)       

Of which (2)

<TABLE border="1" borderColor="#000000" width="_0003"><TBODY><tr><td width="161" height="10"></td></tr><tr>< Td></td><td><img wi="231" he="77" file="02_image005.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(3)       

Since the standard white in s (standard, standard) RGB space is the reference white light under the D65 light source, the D65 light source is the most commonly used artificial daylight in the standard light source, which is different from the spectrum measured by the spectrometer under the endoscope light source. Refer to white light, so these RGB values need to be corrected by color adaptation. In order to be able to accurately estimate the spectral values of the patches, endoscope correction is necessary. Similarly, the spectrum measured by the spectrometer is also converted to the tristimulus values X, Y, Z under the CIE specification by equations (4) to (7) below, where S(λ) is the endoscope Source spectrum, R(λ) is the spectral value of each patch, and , as well as Is a color matching function.

<TABLE border="1" borderColor="#000000" width="_0005"><TBODY><tr><td width="205" height="16"></td></tr><tr>< Td></td><td><img wi="183" he="58" file="02_image013.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(4)       

<TABLE border="1" borderColor="#000000" width="_0006"><TBODY><tr><td width="205" height="19"></td></tr><tr>< Td></td><td><img wi="182" he="58" file="02_image015.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(5)       

<TABLE border="1" borderColor="#000000" width="_0007"><TBODY><tr><td width="204" height="14"></td></tr><tr>< Td></td><td><img wi="183" he="58" file="02_image017.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(6)       

Of which (7)

Through the calculation of equations (4) to (7), the 24-color block spectrum is converted into XYZ value. After color-adaptive conversion, a new XYZ value can be obtained, the XYZ value is converted into RGB value, and the RGB value is set as a matrix [ A], through the third-order polynomial regression of RGB, the conversion relationship between the spectrometer and the endoscope can be found. The following is a matrix of third-order polynomial regression.

<TABLE border="1" borderColor="#000000" width="_0009"><TBODY><tr><td width="204" height="17"></td></tr><tr>< Td></td><td><img wi="170" he="29" file="02_image021.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(8)       

among them (9)

Where "R", "G", and "B" are the RGB values corresponding to each color block taken by the endoscope. The RGB correction of the color block is converted to the tristimulus value XYZ of the CIE standard, which is set to [β]. Finally, the conversion matrix [M] of the endoscope and spectrometer can be obtained from:

<TABLE border="1" borderColor="#000000" width="_0011"><TBODY><tr><td width="216" height="12"></td></tr><tr>< Td></td><td><img wi="158" he="27" file="02_image025.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(10)       

For each pixel in the first pathological image taken by the endoscope, the linear regression matrix [C] can be obtained by multiplication of RGB and the corresponding XYZ value can be obtained by calculation of equations (1) to (3). . The spectrum of each patch (Spectra) (bands from 380 nm to 780 nm) can be obtained by the following equation:

<TABLE border="1" borderColor="#000000" width="_0012"><TBODY><tr><td width="217" height="16"></td></tr><tr>< Td></td><td><img wi="240" he="86" file="02_image027.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(11)       

In this step, the spectrometer measures the reflection or penetration spectrum of the object, and can take the left term in equation (11) to calculate the color. Each pixel in an image can be calculated through this calculation. The reproduction of the color image is completed, and the color image is calculated for the measurement of the analog spectrometer.

Therefore, a comparison of the endoscopic analog spectrum of the 24-color block with the actual measured spectrum can be obtained (see Annex 1). In addition, in order to confirm the feasibility of color reproduction, the color of the 24 color patches captured by the actual endoscope is evaluated by the color difference formula to simulate the color of the 24 color patches for color difference calculation. The calculation process of color difference is as follows:

First, convert the tristimulus value XYZ measured by the two instruments into the chromaticity coordinate values (L*, a*, b*) in the space of the CIE1976 specification, where:

<TABLE border="1"borderColor="#000000"width="_0013"><TBODY><tr><tdwidth="228"height="5"></td></tr><tr><Td></td><td><imgwi="135"he="48"file="02_image029.jpg"img-format="jpg"></img></td></tr></TBODY></TABLE> (12)

(13)

<TABLE border="1" borderColor="#000000" width="_0014"><TBODY><tr><td width="228" height="22"></td></tr><tr>< Td></td><td><img wi="201" he="59" file="02_image033.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(14)       

<TABLE border="1" borderColor="#000000" width="_0015"><TBODY><tr><td width="228" height="15"></td></tr><tr>< Td></td><td><img wi="269" he="67" file="02_image035.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(15)       

Next, calculate the Euclid distance of two points in the chromaticity coordinates of the CIE 1976 specification as the chromatic aberration of two points:

<TABLE border="1" borderColor="#000000" width="_0016"><TBODY><tr><td width="227" height="15"></td></tr><tr>< Td></td><td><img wi="267" he="36" file="02_image037.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(16)       

Using the above formula to calculate the color difference values of the 24 color blocks, the average color difference is about 3.14 (see Annex 2). In general terms, when the color difference is less than 4, it is difficult for the human eye to judge the difference. Such a result shows that the above algorithm can accurately achieve color reproduction, and thus can be applied to the color representation of any image.

In step S102, the first pathological image is imported into an image processing module to obtain a plurality of first simulated spectra of the first pathological images. The image processing module can be composed of a computer with software for installing hyperspectral image processing functions, and the software for the hyperspectral image processing function can be written by a programming software (for example, Microsoft Visual Basic, etc.).

According to histopathology, white light and iodine staining images were divided into normal (Normal), metaplastic (Dysplasia), between metaplasia and esophageal cancer (Dysplasia-ECA), and esophageal cancer (ECA). Among them, when an abnormal change occurs and becomes a tumor cell, it is called a metaplastic defect. NBI magnified images are classified into Intraepithelial papillary capillary loop (IPCL) type 4, IPCL-IV, severe degeneration (Severe Dysplasia), IPCL type 5-1, IPCL-V1 severe dysplasia, IPCL Type 5-1, IPCL-V1 Squamous Cell Carcinoma (SCC), IPCL Type 5-3, and esophageal squamous cell carcinoma (IPCL-V3 SCC).

FIG. 2 is a flow chart of steps of processing an image by the image processing module of the present invention. As shown in FIG. 2, in order to automatically circle the IPCL and record the pixel coordinates of the first pathological image, first, in step S201, the first pathological image obtained by the endoscope is converted into gray through an image grayscale conversion module. Level, image gray level can be achieved by a computer with image processing function. Then, in step S202, the image enhancement of the first pathological image is enhanced by using an image enhancement module, and the contrast enhancement of the image can also be achieved by a computer having an image processing function. Then, in step S203, after enhancing the comparison of the first pathological image, the first pathological image is binarized by the image binarization module, and the image binarization module can be calculated by the programming software in the computer. Image binarization. In step S204, the pixel coordinates of the binarized first pathological image are recorded. Then, in step S205, through the thinning process, the first pathological image can be circled according to the thin line around the binarized first pathological image, and in step S206, the recorded pixel coordinates are exported. To a notepad. The computer with the hyperspectral imaging technology software is then used to read the pathological image of the NBI endoscope and simultaneously read the pixel coordinates in the notebook.

Through the above steps, the first pathological image is converted by the hyperspectral imaging technique, and the simulated spectrum of the first pathological image is obtained. The above image processing steps, such as grayscale processing, image contrast enhancement, image thinning processing and image reading, can be achieved by an image processing module of the present invention. The image processing module is implemented by a computer image processing application. The work of image processing, which is well known to those of ordinary skill in the art, will not be described in detail herein.

The average reflectance spectrum of lesions in white light, iodine staining, and NBI endoscopic images can be obtained by hyperspectral imaging technology. However, white light and iodine stained endoscopic images are sampled by a total of 400 pixels of 20×20; a plurality of NBI endoscopes The first pathological image is directly through the automatic circle selection of the IPCL type, so the pixels of the sample are relatively large, but in the course of the experiment, one thousand coordinate points are uniformly obtained, that is, there are one thousand sets of reflection spectra.

It can be seen from Fig. 3 that the difference in reflectance spectrum is in accordance with the degree of lesions (normal, precancerous lesions, cancer), and the spectral reflectance gradually decreases. Generally, the normal esophageal mucosa is whiter than the color of the gastric mucosa when it produces cancer. In the pre-lesion and cancer lesions, the surface of the esophageal mucosa may be uneven due to the formation of depressions or bulging lesions, and the lesions on the mucosa are darker red than the surrounding mucosa, resulting in more severe lesions and spectral reflectance. The lower. In addition, the spectral reflectance corresponding to the band of blue light and green light is significantly lower than that of the red light band, and the spectral reflectance at a wavelength of 530 nm is found to decrease, which is caused by esophageal cancer with vascular proliferation in the mucosal tissue ( Angiogenesis), which supplies the nutrition and oxygen of cancer cells, and the increased hemoglobin absorbs blue and green light. In addition, there is a peak at about 410 nm and 520 nm. More relevant literature needs to be studied to determine its cause. Conversely, the difference in the reflectance spectrum of the iodine-stained endoscopic image is shown in Fig. 4. The difference in spectrum is that the spectrum reflectance gradually increases according to normal, precancerous lesions and cancer, because the normal esophageal mucosa belongs to the squamous epithelium. The cells, the iodine solution reacts with the glycogen in the cells to stain the mucosa brown, and the hepatic sugar is reduced or disappeared after the cancer cells occupy the epithelium. The iodine solution cannot stain the lesion, so the unstained area on the mucosa is The possibility of cancerous lesions is extremely high, and the area is gradually whiter than the surrounding mucosa depending on the severity of the lesion. The trend of the reflected spectrum of NBI endoscopic images is shown in Figure 5. The difference is that IPCL will show a more divergent, distorted, and irregular pattern according to the degree of canceration of the lesion, and the spectral reflectivity will gradually decrease. The average reflection spectrum of the above three kinds of endoscope images is not changed by the influence of the endoscope light source itself, and the spectrum of the endoscope source is as shown in FIG. 6.

The main component analysis can obtain the spectral characteristics of white light, iodine staining and NBI endoscope image, and then the first main component and the second main component are drawn through the Original software through the software as shown in Annex 3, Annex 4 and Annex 5. Component score chart. Attachment 3 is the result of the spectral characteristics of the white light endoscopic image. It can be found that the fall point of the cancer (ECA) is located between the first principal component (FPC) - 1.90~0.25, and the second principal component (Second Principle Component, SPC) - between 0.35 and 0.35; Dysplasia-ECA has a range of about -0.90 <FPC<0.05, -0.30<SPC<0.30; the range of Dysplasia is about - 0.75<FPC<1.00,-0.20<SPC<0.35; Normal is about 0.30<FPC<3.25, -0.75<SPC<1.25, but between metaplasia, metaplasia-cancer, cancer There is overlap and it is regarded as a fuzzy zone, and the spectrum features of the endoscopic image of normal esophageal mucosa have a tendency to diverge, but the difference of normal esophageal mucosa to canceration can still be seen from the right Left trend. The spectral characteristics of the iodine-stained endoscopic image are shown in Annex 4. The range of normal is about -1.90<FPC<-0.90, -0.19<SPC<0.10; the range of Dysplasia is about At -0.80<FPC<0.50, -0.16<SPC<0.28; metaplasia-cancer (Dysplasia-ECA) ranged from -0.40<FPC<1.70, -0.42<SPC<0.56; cancer (ECA) range Then, it is about 0.00<FPC<2.20, -0.74<SPC<0.18. However, in the same way, there is a blur zone between the metaplasia, metaplasia-cancer, and cancer, and the esophageal mucosa is poor. Both cancer and cancerous endoscopic images have a tendency to diverge, but the spectral characteristics are still roughly different. Conversely, the trend is from left to right depending on the normal esophageal mucosa to cancer. The spectral characteristics of the NBI endoscopic image are shown in Annex 5. The IPCL-V3 SCC falls within the range of -1.70<FPC<-0.90, -0.15<SPC<0.02; the IPCL-V1 SCC has a range of approximately -1.40. <FPC<-0.30, -0.02<SPC<0.40; IPCL-V1 Severe Dysplasia has a range of about -0.75<FPC<1.25, -0.68<SPC<-0.01; IPCL-IV is severely ill-posed (Severe Dysplasia) has a range of about -0.60<FPC<1.60, 0.05<SPC<0.31. It can be found that the four spectral features form a small blurring range, and the spectral characteristics of the IPCL pattern are related to the degree of esophageal canceration. The larger the more converging.

Continuing to refer to Figure 1, the image-processed pixel coordinates can be combined with principal component analysis using hyperspectral image recognition cancer lesions to assess the patient's likelihood of staging in each cancer. In step S103 of the present application for applying the hyperspectral image recognition cancer lesion method, a plurality of triangle ranges are defined in the principal component analysis map. And in step S104, it is determined whether a principal component score of a second simulated spectrum of the second pathological image falls within one of the triangle ranges.

The main component is observed from the perspective of easy-to-view data, so the characteristics of each data can be made more visible. The individual sample values seen from the principal component can be referred to as the Principle Component Score. The formula for the principal component score is as follows:

(17)

Where x _1i , x _2i , . . . , x _ni are the corresponding spectral intensity values for the first, second to nth wavelengths; and x _1i , x _2i , . . . , x _ni are the first The average spectral intensity value from the second to the nth wavelength. These coefficients a _j1 , a _j2 , . . . , a _jn are coefficients of the feature vector after the spectrum takes the covariation matrix. According to the theoretical basis of principal component analysis, the first principal component (y ₁ ) accounts for the most information in the original data and can be regarded as a comprehensive indicator. The second principal component (y ₂ ) accounts for the information of the original data and can be used. Classify each group. Therefore, the analog spectrum obtained by the image processing described in the previous section can be used, and the principal component analysis module can be used to draw the main component score map, and the trend of the spectral features of the pathological image can be observed therefrom.

To assess the likelihood of a patient's staging in each cancer, it is possible to determine whether the principal component score of the simulated spectrum (second simulated spectrum) of the patient's endoscopic image (second pathological image) falls within the defined triangle range. Within, it can be discriminated by the method of discriminating the area of the triangle. As shown in Fig. 7, the hyperspectral image system is used to analyze the image spectrum of the obtained endoscope, apply it to the principal component score map, classify the obtained average analog spectrum, and then find the aforementioned triangle range. The main component analysis plots the main component score map with the largest triangle area as the defined triangle range. For example, first select the leftmost point of the x-axis as the reference point, then select two points and use the computer to calculate The triangle area of these three points, and then choose two points to calculate the other triangle area, compare the size of the two triangle areas, through the above-mentioned triangle area comparison method and computer operation can quickly find the largest area in the principal component analysis chart The triangle range is with its three-point coordinates. When you know the coordinates of the three points of the triangle, you can write the coordinate representation of the vector AB and the vector AC. For example, let the three coordinates of the triangle be A(a1, a2), B(b1, b2), C(c1, c2), then the vector AB=(b1-a1, b2-a2), the vector AC=(c1- A1, c2-a2), and the area of △ABC can be expressed as:

<TABLE border="1" borderColor="#000000" width="_0018"><TBODY><tr><td width="141" height="14"></td></tr><tr>< Td></td><td><img wi="216" he="55" file="02_image047.jpg" img-format="jpg"></img></td></tr></ TBODY></TABLE>(18)       

According to the above-mentioned ΔABC area, if X point is given and its coordinate is X(s, t), the area of ΔXAB, ΔXBC, △XAC can be further represented. It can be known that when the X point falls outside the ΔABC, the following conditions will be met:

△XAB+△XBC+△XAC>△ABC (19)

If it falls inside △ABC or on the boundary, the following conditions will be met:

△XAB+△XBC+△XAC=△ABC (20)

According to the method of discriminating the area of the triangle, it can be known whether the main component score of the simulated spectrum of the patient's endoscope image falls within the above-defined triangle range, and falls on the IPCL-IV Severe Dysplasia, IPCL- V1 is severely deficient (Severe Dysplasia), IPCL-V1 SCC, and IPCL-V3 SCC which is within the triangle range of the staging.

Finally, in step S105, when the main component score of the second analog spectrum falls within one of the triangles, it is confirmed that the second pathological image belongs to a cancer lesion image. The four different degrees of canceration are identified as follows:

The main component score is set as a point to be measured, and if the point to be measured falls within the range of the triangle, the pathological image can be determined as the pathological period of the cancer. For example, esophageal cancer stage 1~4:

1. IPCL-V3 SCC esophageal squamous cell carcinoma: Set U point as the point to be tested. If the following conditions are met, △UAB+△UBC+△UAC=△ABC, the patient will stage the esophageal cancer.

2. IPCL-V1 SCC esophageal squamous cell carcinoma: Set U point as the point to be tested. If the following conditions are met, △UDE+△UEF+△UDF=△DEF, the patient will stage the cancer.

3. IPCL-V1 severely degenerate IPCL type 5-1, severe malformation (IPCL-V1 Severe Dysplasia), set U point as the point to be tested, if the following conditions are met, △UGH+△UHI+△UGI=△GHI The patient is staged for this cancer.

4. IPCL-IV Severe Dysplasia, severe dysplasia, set U point as the point to be tested, if the following conditions are met: △ UJK + △ UKL + △ UJL = △ JKL, the patient staging this cancer.

5. If the U point does not satisfy the above conditions, it will not be recognized.

Through the above-mentioned application of hyperspectral image recognition cancer lesions, cancer lesion images can be digitized, and principal component analysis can be used to quickly assess the likelihood of patients in each cancer stage, effectively and quickly improve the diagnosis efficiency of doctors, and help The patient is treated early.

no

FIG. 1 is a flow chart of an embodiment of a method for identifying a cancer lesion using a hyperspectral image. FIG. 2 is a flow chart of steps of processing an image by the image processing module of the present invention. 3, 4 and 5 are waveform diagrams showing the average result of the reflected spectrum. Figure 6 is a spectrum diagram of the endoscope source. Figure 7 shows a triangle range map for the main component analysis. Figure 8 is a graphical representation of the comparison of the signal intensity ratios of normal and cancerous SHG and TPEF. Figure 9 is a graph comparing the intensity of the infiltration of silver nanoparticles into the esophageal tissue and the intensity of the Raman signal without instillation. Figure 10 is a comparison of the Raman mean spectrum of normal esophageal tissue and carcinogenesis. Figure 11 is a schematic diagram of the main component analysis spindle.

Claims (8)

A method for identifying a cancer lesion by using a hyperspectral image, comprising the steps of: obtaining a plurality of first pathological images from an endoscope, wherein the plurality of first pathological images are a plurality of cancer lesion images; and the first pathological images are merged into one The image processing module obtains a plurality of first simulated spectra of the first pathological images, and draws a principal component score map according to the first simulated spectra; and according to the first simulated spectra, the primary component score map Defining a plurality of triangle ranges; determining whether a principal component score of a second simulated spectrum of a second pathological image falls within one of the triangle ranges; and when the principal component score of the second analog spectrum falls into one of Within the range of the triangles, it is confirmed that the second pathological image belongs to some of the cancer lesion images. The method for identifying a cancer lesion by applying a hyperspectral image according to claim 1, wherein the first pathological image and the second pathological image are transmitted through a hyperspectral image system to obtain the first analog spectrum and the second analog spectrum. The method for applying a hyperspectral image recognition cancer lesion according to claim 1, wherein the first pathological image is imported into the image processing module to obtain the first simulated spectrum of the first pathological images, and according to the The step of drawing the main component score map of the first analog spectrum further includes: converting the first pathological images into gray scales through an image gray scale conversion module; using the image enhancement module to perform the first pathology Image contrast enhancement; binarizing the first pathological images through an image binarization module; recording a plurality of pixel coordinates of the first pathological images after binarization; The pixel coordinates of the pathological image are remitted to draw the main component score map. The method for applying a hyperspectral image recognition cancer lesion according to claim 1, wherein the step of defining a plurality of triangle ranges in the main component score map is to apply a principal component analysis method to draw the main component score map, and then determine the Whether the main component score of the second simulated spectrum of the second pathological image falls within one of the triangles. The method for applying a hyperspectral image recognition cancer lesion according to claim 1, wherein when the main component score of the second simulated spectrum falls within one of the triangles, confirming that the second pathological image belongs to a certain of the cancers In the step of the lesion image, the main component score is a point to be measured, and when the point to be measured is located in a range of one of the triangles, the second pathological image may be determined to correspond to one of the triangles. Cancer pathology. The method for identifying a cancer lesion by applying a hyperspectral image according to claim 1, wherein the plurality of first pathological images are images of a plurality of esophageal cancer lesions. The method for identifying a cancer lesion by applying a hyperspectral image according to claim 1, wherein the first pathological image and the second pathological image are images of blood vessels in the epidermis. The method for applying a hyperspectral image recognition cancer lesion according to claim 1, wherein in the step of defining a plurality of triangle ranges in the principal component score map, the triangle ranges are found according to the first simulated spectrum. The first analog spectrum has the largest triangle area in the principal component score map as the triangle ranges.