Respiratory tract tumor cell intelligent identification system and method based on deep neural network

Abstract

The invention belongs to the technical field of medical image analysis, and discloses a respiratory tract tumor cell intelligent identification system and method based on a deep neural network, and the system comprises a data collection module which collects cell image data of respiratory tract tumor cells and corresponding image description labels; the image processing module is used for carrying out adaptive energy compensation on the cell image data to obtain a pixel compensation graph, and carrying out multi-modal enhancement on the pixel compensation graph to obtain a detail enhancement graph; the space enhancement module is used for carrying out structure explicit on the detail enhancement graph to obtain a cell explicit graph; the model construction module is used for constructing a tumor cell recognition model based on the cell explicit image and the corresponding image description tag, and realizing high-precision recognition of the tumor cells based on the constructed tumor cell recognition model; and the efficiency and the accuracy of identifying the respiratory tumor cells are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and more specifically, to an intelligent recognition system and method for respiratory tract tumor cells based on a deep neural network. Background Art

[0002] With the continuous development of medical imaging, molecular biology, and artificial intelligence technologies, tumor diagnosis, especially the early detection of respiratory tract tumors, has become an important research direction in the field of clinical medicine; for respiratory tract tumors, especially lung cancer, tracheal cancer, laryngeal cancer, etc., early detection is crucial for improving the survival rate of patients; in recent years, with the successful application of artificial intelligence technologies, especially deep learning algorithms, in the medical field, medical image analysis technology has made remarkable progress, especially in the early diagnosis of cancer, cytological analysis, and disease prediction, achieving unprecedented results.

[0003] Currently, the diagnosis of respiratory tract tumors mostly relies on means such as imaging examinations, tissue biopsies, and molecular biology detections; traditional medical imaging technologies often have problems such as low image quality, high noise, and serious background interference in tumor cell detection. Especially in low-resolution images, the details of tumor cells may be blurred or blocked. Especially when cells are in contact or adhered to each other, the blurred boundaries make it impossible for segmentation algorithms to accurately identify and distinguish different cells, resulting in difficulties in morphological analysis and further diagnosis; traditional image processing methods, such as threshold segmentation and edge detection, usually rely on manually designed feature extraction algorithms, lacking in-depth understanding and adaptability to cell morphology and complex structures, resulting in limited effects in practical applications. Although deep learning (especially convolutional neural networks, CNNs) has made certain progress in the field of medical image processing in recent years, most existing deep learning models mainly rely on fixed network structures, lacking full consideration of image quality, noise interference, and cell dynamic changes, and still having problems of insufficient accuracy and robustness; in many tumor cells, the chromatin distribution in the cell nucleus is non-uniform, which makes the morphological and structural analysis of the cell nucleus more complex. The variation in chromatin distribution can affect the discrimination of tumors. Especially in the early stage of malignant tumors, the changes in chromatin are not always obvious, and traditional methods are difficult to identify these subtle differences, thereby causing misidentification of tumor cells.

[0004] In view of this, the present invention proposes an intelligent recognition system and method for respiratory tract tumor cells based on a deep neural network to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention provides the following technical solution: An intelligent recognition system for respiratory tract tumor cells based on a deep neural network, comprising: Data acquisition module: Collect cell image data of respiratory tumor cells and corresponding image description tags; Image processing module: Includes a compensation unit and an enhancement unit. The compensation unit performs adaptive energy compensation on the cell image data to obtain a pixel compensation map; the enhancement unit performs multi-modal enhancement on the pixel compensation map to obtain a detail enhancement map; Spatial enhancement module: Explicitly shows the structure of the detail enhancement map to obtain a cell visualization map; Model construction module: Build a tumor cell recognition model based on the cell visualization map and the corresponding image description tags, and achieve high-precision recognition of tumor cells based on the constructed tumor cell recognition model.

[0006] Furthermore, the acquisition method of the cell image data includes: Labeling tumor cells by fluorescence labeling method, imaging the labeled cells with a fluorescence microscope to obtain cell image data.

[0007] Furthermore, the method for performing adaptive energy compensation on the cell image data includes: Grayscale the cell image data using the RGB weighted average method to obtain a cell grayscale map; Based on the pixel width and pixel height of the cell grayscale map, mark the coordinates of each pixel with the pixel at the lower left corner of the cell grayscale map as the pixel origin to obtain the pixel coordinates of each pixel; Preset a compensation window. For pixel A in the cell grayscale map, use pixel A as the center of the compensation window, and use the other pixels enclosed within the compensation window as the window pixel set of pixel A; Perform equalization definition on each pixel in the cell grayscale map based on the window pixel set to obtain an equalization coefficient; Preset an equalization threshold, and use the pixels with an equalization coefficient greater than the equalization threshold as detail pixel points. Based on the window pixel set of the detail pixel points, calculate the grayscale probability of each pixel in the window pixel set, which is the number of each pixel grayscale in the window pixel set divided by the size of the pixel grayscale set; Perform pixel energy compensation on the detail pixel points based on the grayscale probability to obtain a compensated grayscale; Perform grayscale compensation on the detail pixel points based on the compensated grayscale, and adjust the pixel grayscale of the detail pixel points to the corresponding compensated grayscale to obtain a pixel compensation map.

[0008] Furthermore, the calculation formula for performing equalization definition on each pixel in the cell grayscale map is: ; where represents the equalization coefficient of the pixel at the pixel coordinate , represents the size of the window pixel set, represents the pixel grayscale of the th pixel in the window pixel set, represents the pixel grayscale of the th pixel in the window pixel set, represents the horizontal axis value of the pixel coordinate, The vertical axis value representing the pixel coordinates; The formula for compensating the pixel energy of the detailed pixel points is: ; where represents the compensated gray level of the pixel at the pixel coordinate ; represents the pixel gray level of the pixel at the pixel coordinate ; represents the energy parameter; represents the gray level probability of the th pixel in the window pixel set; represents the pixel gray level of the th pixel in the window pixel set.

[0009] Furthermore, the method for multi-modal enhancement of the pixel compensation map includes: Performing local analysis on each pixel in the pixel compensation map using a standard rate-of-change operator to obtain the pixel rate of change; constructing a description vector for each pixel based on the pixel rate of change and the pixel gray level; for pixel A in the pixel compensation map, taking the pixels in the upper, lower, left, and right four directions adjacent to pixel A as the adjacent pixels of pixel A, and all the adjacent pixels of pixel A form the adjacent pixel set of pixel A; taking each pixel in the pixel compensation map as a point to be evaluated, and performing edge evaluation on the point to be evaluated based on the adjacent pixel set to obtain the boundary evaluation value; Presetting an evaluation threshold, taking the point to be evaluated with a boundary evaluation value greater than or equal to the evaluation threshold as a boundary point, and performing local change adjustment on the boundary point. The formula for performing local change adjustment on the boundary point is: ; where represents the enhanced rate of change of the boundary point; represents the pixel rate of change of the boundary point; represents the enhancement parameter; represents the size of the adjacent pixel set of the boundary point; represents the pixel rate of change of the th adjacent pixel in the adjacent pixel set of the boundary point; represents the similarity evaluation function; represents the description vector of the th adjacent pixel in the adjacent pixel set of the boundary point; represents the description vector of the boundary point; performing the inverse operation of local analysis on the boundary point based on the enhanced rate of change to obtain the boundary gray level, and adjusting the pixel gray level of the boundary point to the corresponding boundary gray level based on the boundary gray level to obtain the detailed enhancement map.

[0010] Furthermore, the formula for performing edge evaluation on the point to be evaluated is: ; among them, represents the boundary evaluation value, represents the size of the adjacent pixel set, represents the feature category of the description vector, represents the edge parameter, represents the -th adjacent pixel in the adjacent pixel set of the point to be evaluated, and the value of the -th feature in the description vector of the adjacent pixel, represents the value of the -th feature in the description vector of the point to be evaluated.

[0011] Furthermore, the method for structure visualization of the detail enhancement map includes: Taking the pixel at the center point of the detail enhancement map as the pixel origin to construct a pixel coordinate system, using the vertical pixel distance and horizontal pixel distance of other pixels from the pixel origin as the mapping position, and using the pixel gray level as the attribute, mapping the detail enhancement map to the pixel coordinate system; performing image spectrum transformation on the pixel coordinate system to obtain the image spectrum; for the image spectrum, recording the spectrum amplitude and spectrum phase at each spectrum level component scale of the image spectrum, and performing spectrum consistent enhancement on the corresponding pixels in the detail enhancement map based on the spectrum amplitude and spectrum phase. The formula for performing spectrum consistent enhancement on the corresponding pixels in the detail enhancement map is: ; among them, represents the enhanced spectrum of the corresponding pixel at the frequency domain coordinate , represents the spectrum amplitude, represents the spectrum phase, represents pi, represents the sum of the image spectrum values traversed from the starting frequency domain coordinate of the image spectrum to the frequency domain coordinate ; constructing a spectrum inverse function based on the formula of the image spectrum transformation, calculating the enhanced spectrum through the spectrum inverse function to obtain the spectrum enhanced gray level value, and adjusting the image gray level based on the spectrum enhanced gray level value to make it match the corresponding spectrum enhanced gray level value to obtain the cell visualization map.

[0012] Furthermore, the formula for performing image spectrum transformation on the mapped coordinate system is: ; among them, represents the image spectrum at the frequency domain coordinate , represents the pixel gray level of the pixel at the coordinate in the pixel coordinate system, represents the imaginary unit, represents pi, represents half of the pixel height in the detail enhancement map, Represents half of the pixel width in the detailed enhancement map, Represents the horizontal axis value of the pixel in the pixel coordinate system, Represents the vertical axis value of the pixel in the pixel coordinate system, Represents integrating the horizontal axis value, Represents integrating the vertical axis value, Represents the spectral horizontal component corresponding to the horizontal axis value, Represents the spectral vertical component corresponding to the vertical axis value.

[0013] Furthermore, the construction method of the tumor cell recognition model includes: Based on the cell manifestation map and the corresponding image description label, using the CNN model as the initial model of the tumor cell recognition model, using the cell manifestation map and the corresponding image description label as the training data, using the training data as the training sample set, training the CNN model with the training sample set, using the cell manifestation map and the corresponding image description label as the input data of the tumor cell recognition model, and using the predicted image prediction label as the output data of the tumor cell recognition model; taking minimizing the error between the actual image description label and the image prediction label predicted by the tumor cell recognition model as the training objective, taking the recall rate function as the loss function of the tumor cell recognition model, and stopping training to obtain the tumor cell recognition model when the loss function converges.

[0014] An intelligent recognition method for respiratory tract tumor cells based on a deep neural network, including: S1. Collect cell image data of respiratory tract tumor cells and the corresponding image description labels; S2. Perform adaptive energy compensation on the cell image data to obtain a pixel compensation map; perform multi-modal enhancement on the pixel compensation map to obtain a detailed enhancement map; S3. Perform structure manifestation on the detailed enhancement map to obtain a cell manifestation map; S4. Construct a tumor cell recognition model based on the cell manifestation map and the corresponding image description labels, and achieve high-precision recognition of tumor cells based on the constructed tumor cell recognition model. The technical effects and advantages of the intelligent recognition system and method for respiratory tract tumor cells based on a deep neural network of the present invention: By performing adaptive energy compensation on cell image data, the present invention improves the overall quality of cell images, better displays the contours and details of cells, and provides a clearer view especially of important structures such as cell nuclei and chromatin. By performing multi-modal enhancement on the pixel compensation map, the edge information in the image is effectively enhanced, making the distinction between tumor cells and surrounding normal cells or the background more obvious, significantly improving the clarity of cell boundaries, eliminating blurred areas, and effectively enhancing the accuracy of accurate identification of tumor cells. By performing structure manifestation on the detail enhancement map, the subtle structural differences between cancer cells and normal cells in morphology can be effectively strengthened, making key information such as cell nuclei and chromatin more prominent in the image, providing higher accuracy for automated cell recognition systems, and also providing clearer reference images for manual inspection, thereby improving the reliability of tumor cell recognition. By constructing a tumor cell recognition model, more reliable data support is provided for pathological diagnosis, treatment plan formulation, etc., greatly improving the efficiency and accuracy of tumor cell recognition and reducing the occurrence of human errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of an intelligent respiratory tract tumor cell recognition system based on a deep neural network according to the present invention; Figure 2 It is a schematic diagram of an intelligent respiratory tract tumor cell recognition method based on a deep neural network according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] Embodiment 1; Please refer to Figure 1 As shown, the intelligent respiratory tract tumor cell recognition system based on a deep neural network in this embodiment includes: Data acquisition module: Acquire cell image data of respiratory tract tumor cells and corresponding image description tags; Image processing module: It includes a compensation unit and an enhancement unit. The compensation unit performs adaptive energy compensation on the cell image data to obtain a pixel compensation map; the enhancement unit performs multi-modal enhancement on the pixel compensation map to obtain a detail enhancement map; Spatial enhancement module: Perform structure manifestation on the detail enhancement map to obtain a cell manifestation map; Model construction module: Construct a tumor cell recognition model based on the cell manifestation map and the corresponding image description tags, and achieve high-precision recognition of tumor cells based on the constructed tumor cell recognition model; Each module is connected in a wired and / or wireless manner to achieve data transmission between modules; The cell image data is obtained by labeling tumor cells using the fluorescence labeling method and performing cell imaging on the labeled cells using a fluorescence microscope. The image description tags are the annotation information of the tumor cells corresponding to the cell image data.

[0018] The methods for adaptive energy compensation of cell image data include: Use the RGB weighted average method to grayscale the cell image data to obtain a cell grayscale image; based on the pixel width and pixel height of the cell grayscale image, mark the coordinates of each pixel with the pixel at the lower left corner of the cell grayscale image as the pixel origin to obtain the pixel coordinates of each pixel; preset a compensation window. For pixel A in the cell grayscale image, use pixel A as the center of the compensation window, and use the other pixels enclosed within the compensation window as the window pixel set of pixel A; perform equalization definition on each pixel in the cell grayscale image based on the window pixel set. The calculation formula for performing equalization definition on each pixel in the cell grayscale image is: ; where, represents the equalization coefficient of the pixel at the pixel coordinate of ; represents the size of the window pixel set, represents the th pixel in the window pixel set, represents the th pixel in the window pixel set, represents the horizontal axis value of the pixel coordinate, represents the vertical axis value of the pixel coordinate; the smaller the equalization coefficient, the simpler the area where the pixel is located, and the more likely it is the background area. The larger the equalization coefficient, the more complex the area where the pixel is located, containing more textures and details; Preset an equalization threshold, which is set by those skilled in the art based on the actual situation; use the pixels with equalization coefficients greater than the equalization threshold as detail pixel points, and calculate the gray probability of each pixel in the window pixel set of the detail pixel points, by dividing the number of each pixel gray level in the window pixel set by the size of the pixel gray level set; perform pixel energy compensation on the detail pixel points based on the gray probability. The formula for performing pixel energy compensation on the detail pixel points is: ; where, represents the compensated gray level of the pixel at the pixel coordinate of ; represents the pixel coordinate of The pixel gray level of the pixel at the position, represents an energy parameter for adjusting the sensitivity of energy compensation, represents the gray level probability of the th pixel in the window pixel set, represents the pixel gray level of the th pixel in the window pixel set; based on the compensated gray level, gray level compensation is performed on the detail pixel points, and the pixel gray level of the detail pixel points is adjusted to the corresponding compensated gray level to obtain a pixel compensation map; By performing adaptive energy compensation on the cell image data, the brightness of each region is reasonably increased or suppressed, avoiding overcompensation or uneven illumination, enhancing the details in the image, thereby improving the overall quality of the image. Especially in cell morphology analysis, it can better display the cell contour.

[0019] The methods for multi-modal enhancement of the pixel compensation map include: Performing local analysis on each pixel in the pixel compensation map using a standard rate-of-change operator (performing gradient operation on the pixel using the rate-of-change operator) to obtain the pixel rate of change. Common rate-of-change operators include the Sobel operator and the Prewitt operator; constructing a description vector for each pixel based on the pixel rate of change and the pixel gray level; for pixel A in the pixel compensation map, using the pixels in the four adjacent directions of up, down, left, and right to pixel A as the adjacent pixels of pixel A, and all the adjacent pixels of pixel A form the adjacent pixel set of pixel A; taking each pixel in the pixel compensation map as a point to be evaluated, and performing edge evaluation on the point to be evaluated based on the adjacent pixel set. The formula for performing edge evaluation on the point to be evaluated is: ; where, represents the boundary evaluation value, represents the size of the adjacent pixel set, represents the feature category of the description vector, represents an edge parameter for adjusting the scale of the boundary evaluation value, represents the th adjacent pixel in the adjacent pixel set of the point to be evaluated, and the th feature value in the description vector of the th adjacent pixel, represents the th feature value in the description vector of the point to be evaluated; A preset evaluation threshold, which is set by those skilled in the art based on the actual situation. Taking the points to be evaluated with the boundary evaluation value greater than or equal to the evaluation threshold as boundary points, and performing local change adjustment on the boundary points. The formula for performing local change adjustment on the boundary points is: ; where, represents the enhanced rate of change of the boundary point, The pixel change rate representing the boundary point, represents the enhancement parameter used to adjust the balance between the pixel change rate of the boundary point and the pixel change rates of the adjacent pixels in the adjacent pixel set, represents the size of the adjacent pixel set of the boundary point, represents the th adjacent pixel's pixel change rate in the adjacent pixel set of the boundary point, represents the similarity evaluation function. Common similarity evaluation function formulas include the cosine similarity function and the Euclidean distance function, represents the th adjacent pixel's description vector in the adjacent pixel set of the boundary point, represents the description vector of the boundary point; the inverse operation of local analysis of the boundary point based on the enhanced change rate (the reverse operation of local analysis of the boundary point using the change rate operator) to obtain the boundary gray level, and adjust the pixel gray level of the boundary point to the corresponding boundary gray level based on the boundary gray level to obtain the detail enhancement map; By locally adjusting the boundary points, the boundary details of the pixel compensation map are significantly enhanced, effectively improving the clarity of the cell boundary in the pixel compensation map, eliminating the unclear boundary influence caused by blurring, and being able to effectively improve the effects of cell segmentation and feature extraction.

[0020] The ways to make the structure of the detail enhancement map explicit include: Construct a pixel coordinate system with the pixel at the center point of the detail enhancement map as the pixel origin, use the vertical pixel distance and horizontal pixel distance of other pixels from the pixel origin as the mapping positions, and use the pixel gray level as the attribute to map the detail enhancement map to the pixel coordinate system; perform image spectrum transformation on the pixel coordinate system. The formula for performing image spectrum transformation on the mapped coordinate system is: ; where, represents the image spectrum at the frequency domain coordinate of , represents the pixel gray level of the pixel at the coordinate of in the pixel coordinate system, represents the imaginary unit, represents the pi, represents half of the pixel height in the detail enhancement map, represents half of the pixel width in the detail enhancement map, represents the horizontal axis value of the pixel in the pixel coordinate system, represents the vertical axis value of the pixel in the pixel coordinate system, represents the integration of the horizontal axis value, represents the integration of the vertical axis value, represents the spectral horizontal component corresponding to the horizontal axis value, Represents the vertical component of the spectrum corresponding to the value on the vertical axis; for the image spectrum, record the spectrum amplitude and spectrum phase of the image spectrum at each scale of the spectrum horizontal component. Based on the spectrum amplitude and spectrum phase, perform spectrum-consistent enhancement on the corresponding pixels in the detail-enhanced image. The formula for spectrum-consistent enhancement is: ; where represents the enhanced spectrum of the corresponding pixel at the frequency domain coordinate , represents the spectrum amplitude, represents the spectrum phase, represents pi, represents the sum of the image spectrum values traversed from the starting frequency domain coordinate of the image spectrum to the frequency domain coordinate ; Construct a spectrum inverse function based on the formula for image spectrum transformation (i.e., a function to reverse-infer pixel grayscale from the image spectrum). Calculate the enhanced spectrum through the spectrum inverse function to obtain the spectrum-enhanced grayscale value. Based on the spectrum-enhanced grayscale value, adjust the image grayscale to match the corresponding spectrum-enhanced grayscale value to obtain the cell visualization map; In tumor cell recognition, the slight morphological differences between cancer cells and normal cells (such as abnormal nuclear morphology and uneven chromatin distribution) are key diagnostic bases. Through spectrum-consistent enhancement operations, it can effectively strengthen the subtle structural differences, improve the effective information in the image (such as the texture of cell nuclei and chromatin), contribute to an automated cell recognition system, and also provide a more accurate reference basis for manual inspection, thereby improving the accuracy and reliability of diagnosis.

[0021] The construction method of the tumor cell recognition model includes: Based on the cell visualization map and the corresponding image description labels, use the CNN model as the initial model of the tumor cell recognition model. Use the cell visualization map and the corresponding image description labels as training data, and use the training data as the training sample set. Use the training sample set to train the CNN model. Use the cell visualization map and the corresponding image description labels as the input data of the tumor cell recognition model, and use the predicted image prediction label as the output data of the tumor cell recognition model; Take minimizing the error between the actual image description label and the image prediction label predicted by the tumor cell recognition model as the training objective, and use the recall rate function as the loss function of the tumor cell recognition model. When the loss function converges, stop training to obtain the tumor cell recognition model; Based on the constructed tumor cell recognition model, achieve high-precision recognition of tumor cells.

[0022] In this embodiment, by performing adaptive energy compensation on the cell image data, the overall quality of the cell image is improved, the contours and details of the cells are better displayed, and a clearer view is provided especially for important structures such as the cell nucleus and chromatin. By performing multi-modal enhancement on the pixel compensation map, the edge information in the image is effectively enhanced, making the difference between tumor cells and surrounding normal cells or the background more obvious, significantly improving the clarity of the cell boundary, eliminating blurred areas, and effectively enhancing the accuracy of accurate identification of tumor cells. By performing structure visualization on the detail enhancement map, the subtle structural differences between cancer cells and normal cells in morphology can be effectively strengthened, making key information such as the cell nucleus and chromatin more prominent in the image, providing higher accuracy for an automated cell recognition system, and also providing a clearer reference image for manual inspection, thereby improving the reliability of tumor cell recognition. By constructing a tumor cell recognition model, more reliable data support is provided for pathological diagnosis, treatment plan formulation, etc., greatly improving the efficiency and accuracy of tumor cell recognition and reducing the occurrence of human errors.

[0023] Embodiment 2; Please refer to Figure 2 As shown, for the parts not described in detail in this embodiment, refer to the description in Embodiment 1. The intelligent recognition method for respiratory tract tumor cells based on a deep neural network includes: S1. Collect cell image data of respiratory tract tumor cells and corresponding image description labels; S2. Perform adaptive energy compensation on the cell image data to obtain a pixel compensation map; perform multi-modal enhancement on the pixel compensation map to obtain a detail enhancement map; S3. Perform structure visualization on the detail enhancement map to obtain a cell visualization map; S4. Construct a tumor cell recognition model based on the cell visualization map and the corresponding image description labels, and achieve high-precision recognition of tumor cells based on the constructed tumor cell recognition model.

[0024] Embodiment 3; This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the running mode of the above-provided intelligent recognition method for respiratory tract tumor cells based on a deep neural network.

[0025] Since the electronic device introduced in this embodiment is the electronic device used to implement the intelligent recognition method of respiratory tract tumor cells based on a deep neural network in the embodiments of the present application, based on the intelligent recognition method of respiratory tract tumor cells based on a deep neural network introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manners and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device used in the intelligent recognition method of respiratory tract tumor cells based on a deep neural network in the embodiments of the present application, it falls within the scope of protection of the present application.

[0026] The above formulas are all dimensionless and take their numerical values for calculation. The formula is obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold selection in the formula are set by those skilled in the art according to the actual situation.

[0027] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for ordinary technical users in the technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. Intelligent identification system of respiratory tumor cells based on deep neural network, characterized by: include: Data acquisition module: collects cell image data of respiratory tumor cells and corresponding image description labels; Image processing module: including a compensation unit and an enhancement unit. The compensation unit performs adaptive energy compensation on the cell image data to obtain a pixel compensation map. The enhancement unit performs multi-modal enhancement on the pixel compensation image to obtain a detail enhancement image; Spatial enhancement module: structurally visualize the detail enhancement map to obtain the cell visualization map; Model building module: Build a tumor cell recognition model based on the cell visualization map and the corresponding image description labels, and achieve high-precision recognition of tumor cells based on the constructed tumor cell recognition model.

2. The deep neural network-based intelligent identification system for respiratory tumor cells according to claim 1 is characterized in that: The method for acquiring the cell image data includes: marking tumor cells by a fluorescent labeling method, and performing cell imaging on the marked cells using a fluorescent microscope to obtain the cell image data.

3. The intelligent identification system for respiratory tumor cells based on deep neural network according to claim 2 is characterized in that: The method of performing adaptive energy compensation on cell image data includes: The RGB weighted average method is used to grayscale the cell image data to obtain a cell grayscale image; based on the pixel width and pixel height of the cell grayscale image, the pixel in the lower left corner of the cell grayscale image is used as the pixel origin to mark the coordinates of each pixel to obtain the pixel coordinates of each pixel; a compensation window is preset, and for pixel A in the cell grayscale image, pixel A is used as the center of the compensation window, and other pixels circled in the compensation window are used as the window pixel set of pixel A; each pixel in the cell grayscale image is balanced based on the window pixel set to obtain a balance coefficient; a balance threshold is preset, and pixels with a balance coefficient greater than the balance threshold are used as detail pixels. Based on the window pixel set of the detail pixels, the grayscale probability of each pixel in the window pixel set is calculated, and the number of each pixel grayscale in the window pixel set is divided by the size of the pixel grayscale set; pixel energy compensation is performed on the detail pixels based on the grayscale probability to obtain a compensated grayscale; grayscale compensation is performed on the detail pixels based on the compensated grayscale, and the pixel grayscale of the detail pixels is adjusted to the corresponding compensated grayscale to obtain a pixel compensation image.

4. The intelligent identification system for respiratory tumor cells based on deep neural network according to claim 3 is characterized in that: The calculation formula for balancing and defining each pixel in the cell grayscale image is: ;in, The pixel coordinates are The equalization coefficient of the pixel at Represents the size of the window pixel set, Represents the first pixel in the window The pixel grayscale of pixels, Represents the first pixel in the window The pixel grayscale of pixels, The horizontal axis value represents the pixel coordinate. The vertical axis value representing the pixel coordinate; The formula for performing pixel energy compensation on detail pixels is: ;in, The pixel coordinates are The compensated gray level of the pixel at The pixel coordinates are The pixel grayscale of the pixel at represents the energy parameter, Represents the first pixel in the window The gray level probability of a pixel, Represents the first pixel in the window The pixel grayscale of pixels.

5. The deep neural network-based intelligent identification system for respiratory tumor cells according to claim 4 is characterized in that: The method of performing multi-modal enhancement on the pixel compensation map includes: A standard rate of change operator is used to perform local analysis on each pixel in the pixel compensation map to obtain the pixel change rate; a description vector of each pixel is formed based on the pixel change rate and the pixel grayscale; for pixel A in the pixel compensation map, pixels in the four directions of up, down, left and right adjacent to pixel A are used as adjacent pixels of pixel A, and all adjacent pixels of pixel A constitute the adjacent pixel set of pixel A; each pixel in the pixel compensation map is used as a point to be evaluated, and an edge evaluation is performed on the point to be evaluated based on the adjacent pixel set to obtain a boundary evaluation value; The evaluation threshold is preset, and the points to be evaluated whose boundary evaluation values ​​are greater than or equal to the evaluation threshold are taken as boundary points. The boundary points are adjusted for local changes. The formula for adjusting the boundary points for local changes is: ;in, represents the enhanced change rate of the boundary point, Represents the pixel change rate of the boundary point, represents the enhancement parameter, The size of the set of adjacent pixels representing the boundary points, The first pixel in the adjacent pixel set representing the boundary point The pixel change rate of adjacent pixels, represents the similarity rating function, The first pixel in the adjacent pixel set representing the boundary point The description vector of adjacent pixels, A description vector representing a boundary point; performing an inverse operation of local analysis on the boundary point based on the enhancement change rate to obtain the boundary grayscale; adjusting the pixel grayscale of the boundary point to the corresponding boundary grayscale based on the boundary grayscale to obtain a detail enhancement map.

6. The intelligent identification system for respiratory tumor cells based on deep neural network according to claim 5 is characterized in that: The formula for edge evaluation of the point to be evaluated is: ;in, represents the boundary assessment value, represents the size of the adjacent pixel set, Represents the feature category describing the vector, represents the edge parameter, Represents the neighboring pixel set of the point to be evaluated. The description vector of the neighboring pixels The value of the feature, Represents the first The value of a feature.

7. The deep neural network-based intelligent identification system for respiratory tumor cells according to claim 6 is characterized in that: The method of structurally displaying the detail enhancement map includes: The pixel coordinate system is constructed with the pixel at the center of the detail enhancement image as the pixel origin, the vertical pixel distance and horizontal pixel distance from other pixels to the pixel origin are used as the mapping position, and the pixel grayscale is used as the attribute to map the detail enhancement image to the pixel coordinate system; the pixel coordinate system is transformed into an image spectrum to obtain the image spectrum; for the image spectrum, the spectrum amplitude and spectrum phase of the image spectrum at each spectrum horizontal component scale are recorded, and the corresponding pixels in the detail enhancement image are spectrally enhanced based on the spectrum amplitude and spectrum phase. The formula for spectrally enhancing the corresponding pixels in the detail enhancement image is: ;in, The frequency domain coordinates are The enhanced spectrum of the corresponding pixel at represents the spectrum amplitude, represents the spectral phase, represents pi, Represents the traversal from the starting frequency domain coordinate of the image spectrum to the frequency domain coordinate The image spectrum values ​​at the locations are summed; a spectrum inverse function is constructed based on the formula for image spectrum conversion, the enhanced spectrum is calculated through the spectrum inverse function to obtain the spectrum enhancement grayscale value, and based on the spectrum enhancement grayscale value, the image grayscale is adjusted to match the corresponding spectrum enhancement grayscale value to obtain a cell visualization map.

8. The deep neural network-based intelligent identification system for respiratory tumor cells according to claim 7 is characterized in that: The formula for performing image spectrum conversion on the mapping coordinate system is: ;in, The frequency domain coordinates are The image spectrum at Represents the coordinates in the pixel coordinate system The pixel grayscale of the pixel at represents the imaginary unit, represents pi, Represents half of the pixel height in the detail enhancement image. Represents half the pixel width in the detail enhancement image. Represents the horizontal axis value of the pixel in the pixel coordinate system, Represents the vertical axis value of the pixel in the pixel coordinate system, Represents the integration of the horizontal axis value. Represents the integration of the vertical axis value. Represents the horizontal spectrum component corresponding to the horizontal axis value, Represents the vertical component of the spectrum corresponding to the vertical axis value.

9. The deep neural network-based intelligent identification system for respiratory tumor cells according to claim 8, characterized in that: The method for constructing the tumor cell recognition model includes: Based on the cell visualization image and the corresponding image description label, the CNN model is used as the initial model of the tumor cell recognition model, the cell visualization image and the corresponding image description label are used as training data, and the training data is used as a training sample set. The CNN model is trained using the training sample set, and the cell visualization image and the corresponding image description label are used as input data of the tumor cell recognition model, and the predicted image prediction label is used as output data of the tumor cell recognition model; minimizing the error between the actual image description label and the image prediction label predicted by the tumor cell recognition model is used as the training goal, and the recall rate function is used as the loss function of the tumor cell recognition model. When the loss function reaches convergence, the training is stopped to obtain the tumor cell recognition model.

10. A method for intelligently identifying respiratory tract tumor cells based on a deep neural network, which is implemented based on the intelligent system for intelligently identifying respiratory tract tumor cells based on a deep neural network according to any one of claims 1 to 9, comprising: S1, collecting cell image data of respiratory tumor cells and corresponding image description labels; S2, performing adaptive energy compensation on the cell image data to obtain a pixel compensation map; performing multimodal enhancement on the pixel compensation map to obtain a detail enhancement map; S3, structurally visualize the detail enhancement image to obtain a cell visualization image; S4. A tumor cell recognition model is constructed based on the cell visualization map and the corresponding image description labels, and high-precision recognition of tumor cells is achieved based on the constructed tumor cell recognition model.

Images