Single bacterium spectrum detection method based on filter array

Abstract

The invention relates to a single bacterium spectrum detection method based on a filter array, which is characterized in that a 5 * 5 optical filter array is designed to cover a 614-974nm spectrum range, and a super-resolution demosaicing technology is combined, so that the imaging resolution is remarkably improved. And the reconstructed three-dimensional hyperspectral data is analyzed by using a 3D convolutional neural network model, so that efficient and accurate detection on a single bacterium level is realized. The technology is especially suitable for rapid identification of food-borne bacteria, and the efficiency of public health and food safety detection is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of spectral detection, and in particular relates to a single bacteria spectral detection method based on a filter array. Background Art

[0002] Foodborne pathogens pose a major threat to public health and safety and a challenge to the modern food industry. Bacteria are widely present in various environments and have a huge impact on human life. Beneficial bacteria can promote human health, while pathogenic bacteria may cause harm to humans by contaminating food. These pathogens present in various foods can enter the human body and multiply rapidly, increasing the risk of health crises. Foodborne diseases refer to diseases caused by pathogenic factors such as toxic and harmful substances (including biological pathogens) that enter the human body through ingestion. They can generally be divided into infectious and toxic diseases, including common food poisoning, intestinal infectious diseases, zoonotic infectious diseases, parasitic diseases, and diseases caused by chemical toxic and harmful substances. According to statistics from the World Health Organization, about 600 million cases of foodborne diseases occur each year, of which 420,000 people die. Therefore, the detection and identification of bacteria is very important and necessary. Accurate bacterial detection and identification technology is of great significance to improving disease diagnosis and treatment effects and protecting human health.

[0003] Traditional bacterial detection methods mainly include smear microscopy, isolation and culture, serological identification and molecular biological detection technology. Smear microscopy is a quick and simple method that observes the morphology and structure of bacteria under a microscope, but its sensitivity is low and it cannot accurately identify specific species. Moreover, the results largely depend on the experience and skills of the inspectors. Isolation culture and biochemical reaction are the "gold standard" for bacterial detection and can provide relatively accurate results, but this process is time-consuming and complicated to operate, and it is powerless for some microorganisms that are difficult to culture or cannot be cultured. Serological identification methods such as ELISA are widely used in clinical practice because of their rapid detection, but there is a risk of false positives and it is impossible to detect multiple microorganisms at one time. Molecular biological techniques such as PCR have high sensitivity and specificity and can detect extremely low levels of pathogens, but they are expensive, have high requirements for equipment and technology, and may have false positive problems. These traditional methods have their own advantages and disadvantages, and usually need to be used in combination to obtain more accurate test results, but most of them have problems such as complex operation, long time, high cost and insufficient sensitivity.

[0004] In order to overcome the limitations of traditional bacterial detection methods, it is urgent to develop faster and more effective single bacterial detection technologies. In this context, the development of new technologies for rapid clinical microbial detection has brought about a series of innovative methods that aim to provide more accurate and faster detection methods. Among them, microfluidics combined with optical tweezers technology has shown great potential in the field of pathogenic microorganism detection, and can achieve single-cell separation of foodborne pathogens such as Escherichia coli, Salmonella, Staphylococcus aureus and Bacillus. This technology mainly quantifies the fluorescence properties of bacteria through fluorescent dye labeling, and dynamic fluorescence imaging can provide detailed spatial and temporal information of bacterial activities due to its high sensitivity. However, the reliance on fluorescent dyes also brings a series of problems, including spectral overlap, possible adverse effects on cell activity, autofluorescence, and instability of fluorescent dyes, all of which may lead to distortion of measurement accuracy and limit the application of this method in research and diagnostic environments.

[0005] Hyperspectral imaging technology is mainly divided into two types: scanning and snapshot. When scanning hyperspectral imaging technology collects dynamic scenes with significant spatial or spectral changes, it needs to scan point by point or line by line, and cannot quickly obtain the entire 3D data cube. Therefore, it has obvious shortcomings when detecting small targets such as bacteria. In contrast, snapshot hyperspectral imaging technology can obtain a complete data cube in one exposure cycle without the need for a scanning process, so it is more suitable for fast and real-time bacterial detection. Snapshot hyperspectral imaging technology divides the 3D data cube into multiple 2D data for simultaneous measurement, and then recombines the 2D data into a 3D cube through post-processing. This technology can provide detailed spectral characteristics of bacteria and is suitable for the detection of dynamic scenes and small targets. Among these many existing technologies, filter array-based hyperspectral cameras stand out due to their significant advantages. This type of camera can quickly collect data and capture high-throughput dynamic scenes at one time, which is particularly suitable for monitoring fast-moving objects in microfluidic channels. In addition, by eliminating mechanical scanning, they not only simplify the imaging process, but also greatly reduce system complexity, and become a better solution for label-free single bacterial detection microfluidic technology. However, while snapshot hyperspectral imaging technology can achieve real-time monitoring, it also faces the challenge of the mutual limitation of spectral resolution and spatial resolution. Traditional de-mosaicing methods are susceptible to noise interference, which can seriously damage the spatial resolution of the generated image. In bacterial detection applications, due to the tiny size of bacteria, usually at the micrometer or even nanometer level, the accuracy requirements of the detection technology are extremely high. Ensuring image clarity to avoid any loss of accuracy is the key to achieving reliable detection.

[0006] In view of the shortcomings of the prior art, the present invention provides a hyperspectral imaging technology based on a filter array and its application in bacterial detection. This technology significantly improves the imaging resolution by 8 times through super-resolution de-mosaicing technology. The captured bacterial mosaic image is reconstructed into three-dimensional hyperspectral data, and the data is analyzed using a 3D convolutional neural network model that integrates bacterial structure and spectral features, and the recognition accuracy of pathogenic bacteria is increased to 91.7%. This method greatly improves the efficiency and effectiveness of bacterial detection and provides a promising solution for key applications in public health and the food industry. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a single bacteria spectral detection method based on a filter array. The method is based on the hyperspectral imaging technology of the filter array and is enhanced by super-resolution demosaicing technology to improve the imaging resolution and detection efficiency, thereby realizing rapid and highly specific Raman spectral detection of samples such as bacteria.

[0008] The present invention solves the technical problem by the following technical solutions:

[0009] A single bacteria spectral detection method based on a filter array, characterized by comprising the following steps:

[0010] Design and use a 5 × 5 filter array covering the spectral range of 614 to 974 nm to capture hyperspectral images of bacteria;

[0011] During the detection process, a microscope objective with a magnification of 60 times and a numerical aperture of 1.25 was used to capture microscopic images of bacteria to obtain high-resolution bacterial imaging;

[0012] The bacterial samples are guided to flow through the microfluidic channel in sequence, and a hyperspectral camera based on the filter array is used to continuously image the bacterial samples at a fixed position to quickly capture the dynamic changes of the bacteria;

[0013] The captured bacterial mosaic images are reconstructed into three-dimensional hyperspectral data through super-resolution demosaicing technology to enhance the spatial details of the images and reveal the key structural features of the bacteria;

[0014] A 3D convolutional neural network model that integrates bacterial structure and spectral features is used to analyze the three-dimensional hyperspectral data to achieve efficient and accurate detection at the single bacteria level.

[0015] Moreover, the super-resolution demosaicing technique effectively enhanced the spatial details of bacteria and revealed the key structural features of bacteria by reconstructing sparse data and amplifying the spectral signal by 8 times, which was achieved through weighted convolution, where each missing pixel value was filled by averaging the adjacent known pixel values.

[0016] Moreover, the technology uses bilinear interpolation technology to estimate the missing pixel values ​​and create a preliminary image that emphasizes low-frequency details. Specifically, for the target pixel, its four adjacent pixels in the original image are first determined; then, based on the position of the target pixel relative to these four known pixels, a specific weight calculation formula is used to estimate the value of the target pixel.

[0017] Moreover, a 3D CNN model is constructed to make up for the loss of high-frequency details that may be caused by bilinear interpolation, and the hyperspectral data cube can be efficiently processed to enhance the high-frequency details of the image and improve the accuracy of detection.

[0018] Moreover, the loss function is the weighted sum of the mean square error MSE and the structural similarity index SSIM, with weights w1=0.8 and w2=0.2 respectively. By optimizing this loss function, the model can reduce pixel-level errors and preserve the structural integrity of the image, achieving more accurate and visually consistent reconstruction.

[0019] Moreover, the 3D convolutional neural network model can capture additional spectral information dimensions, providing multiple spectral bands for each pixel, thereby increasing the data dimension while improving the accuracy of bacterial identification. The model starts with a series of convolutional layers configured with 3×3×3 convolutional kernels. While capturing detailed features, the model ensures that the resolution of the input data is fully preserved, avoiding information loss, and extracting richer and more detailed features, providing strong feature support for subsequent bacterial classification and identification.

[0020] The advantages and beneficial effects of the present invention are:

[0021] 1. The snapshot hyperspectral imaging system used in the present invention is equipped with a 5×5 filter array, which can effectively capture hyperspectral data, simulate the imaging process through a neural network, and simultaneously improve the design of the filter and the accuracy of spectral reconstruction, satisfying the characteristics of the filter array's strong spectral modulation capability.

[0022] 2. The present invention establishes a super-resolution demosaicing model to amplify the spectral resolution by 8 times by reconstructing sparse data, effectively enhancing the spatial details of bacteria and revealing key structural features for accurate identification and analysis of bacteria. It not only enhances spectral fidelity, but also significantly improves spatial resolution, providing strong technical support for the application of hyperspectral imaging in the field of bacterial detection.

[0023] 3. The hyperspectral imaging bacteria recognition network based on three-dimensional convolutional neural network (3D CNN) developed by the present invention is different from traditional two-dimensional imaging and can capture additional spectral information dimensions. Each pixel provides multiple spectral bands, thereby providing spectral characteristics of objects carrying detailed bacterial composition information, further improving recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the filter array of the present invention. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0026] A single bacteria spectral detection method based on filter array, the innovation of which is: Figure 1 As shown, the following steps are included:

[0027] ①First, five major foodborne bacterial strains from the BeNa culture collection were selected, including Campylobacter jejuni BNCC360736, Escherichia coli BNCC186347, Listeria monocytogenes BNCC185986, Staphylococcus aureus BNCC376051, and Salmonella typhimurium BNCC365311. Campylobacter jejuni was cultured in Mueller-Hinton broth, while the other strains were cultured aerobically in tryptic soy broth. After cultivation, the bacteria were precipitated by centrifugation and washed with sterile phosphate buffered saline (PBS) to remove residual culture medium and then resuspended in PBS in preparation for spectral analysis. These samples were then introduced into a PDMS-based microfluidic device for subsequent analysis.

[0028] ② Secondly, in order to accurately identify the characteristic spectral bands of foodborne bacteria, the present invention selects 25 different channels, each of which contains near-infrared spectral features that are critical for accurate detection. The filter width of the hyperspectral camera based on the filter array is limited to less than 15nm, and the peak transmittance exceeds 90%, ensuring high-precision detection. The imaging system uses a 5×5 filter array to capture raw data, significantly reducing the spatial density of the raw spectral data to 1 / 25 of its initial volume.

[0029] ③Next, these images are converted into hyperspectral data using a super-resolution demosaicing method. The captured bacterial mosaic images are reconstructed into three-dimensional hyperspectral data. The technique reconstructs sparse data through bilinear interpolation and magnifies it by 8 times to enhance the spatial details of the bacteria. The process first applies bilinear interpolation to a sparse matrix to estimate missing pixels, thereby creating a preliminary image that enhances low-frequency details. Bilinear interpolation is a filter implemented by weighted convolution, in which missing pixel values ​​are filled by averaging adjacent known pixel values. Using a 9×9 filter array, each pixel is convolved with a kernel with a stride of 1. The interpolation weight is inversely proportional to the distance from the center, and the closer the pixel is to the center, the larger the weight. Finally, the selected data enters a dynamic filtering network to enhance image resolution and a residual network to restore high-frequency details of the hyperspectral data. The combination of these networks outputs a restored hyperspectral image with enhanced spectral fidelity and spatial resolution.

[0030] ④ The mosaic image has a size of 512×512 pixels and spans 25 spectral channels. The process first processes H through a weighted convolution filter, followed by a 3×3×33D convolution, a batch normalization (BN) layer, and a rectified linear unit (ReLU) activation. After these initial transformations, the data is divided into two different paths: a dynamic filtering network and a residual network. In the dynamic filtering network, the architecture includes eight dense blocks, each consisting of a 3×3×3 convolution layer, a BN layer and a ReLU activation layer, a 1×1×1 convolution layer, another BN layer, and a ReLU activation layer. After the dense block, the architecture proceeds through a series of convolutional layers designed to generate a refined feature map before the dynamic filter. First, a 3×3 convolution layer is applied to enhance local connectivity patterns and capture more complex features in the processed data. This layer is followed by a ReLU. A 1×1 convolution layer is then used. This layer is mainly used to reduce the dimensionality of the feature map, thereby achieving a more compact feature representation without causing significant information loss. It is followed by another ReLU activation to maintain nonlinearity. This pattern is repeated with an additional 1×1 convolution followed by a ReLU activation. The output of the network is a softmax layer for classification, which converts the refined feature map into a dynamic filter F i ,,n. Each dynamic filter in the network has a size of 5×5. The indexes i, j, and n vary in the range of i∈[1, r×L], j∈[1, r×W], ​​and n∈[1, C], respectively. These dynamic filters F i,j,n Applied to the initial stitching image with M i',j' The corresponding 5×5 area centered at the center. I' and j' represent the bottom surface of i and j divided by r, respectively, effectively mapping each filter to the original image. This mapping process can be seen in the enlarged hyperspectral image Z. i ,,n generates new pixel values.

[0031] ⑤ In order to better align the reconstructed image with the ground truth, the present invention uses a combined loss function, which contains the mean square error (MSE) and the structural similarity index metric (SSIM). The loss function is calculated as the weighted sum of MSE and SSIM, and the formula is as follows:

[0032] L=w1×L mse +w2×L ssim

[0033] SSIM evaluates the visual impact of three features of an image: brightness, contrast, and structure. It compares local patterns of pixel intensities that have been normalized for brightness and contrast. The lower the SSIM value, the greater the degradation of structural information, which corresponds to the lower quality of the reconstructed image. Through testing, the best effect was achieved when the weight of MSE was set to w1=0.8 and the weight of SSIM was set to w2=0.2. The network was trained using the PyTorch framework and optimized for the NVIDIA GeForce RTX 3070 graphics card. The training involved 80,000 iterations with a p size of 8. The step decay learning rate strategy was used to gradually improve the accuracy of the network. The Adam optimizer was used to adjust the network parameters, which was conducive to the effective convergence of the loss function and improved the performance and stability of the model. The settings of the optimizer are β1=0.7, β2=0.997, ε=10 respectively. -8 , to improve the efficiency and accuracy of the training process.

[0034] ⑥ After the above steps, the present invention adopts a three-dimensional convolutional neural network for high-spectral imaging bacteria recognition network to establish a 3D CNN model architecture, which is optimized from five convolution structures using 3×3×3 kernels to capture detailed features across spatial and spectral dimensions while maintaining full data resolution. Each layer uses one step to fully cover the input data, ensuring that all possible feature combinations are learned, followed by convolution, BN standardized output, promoting faster training convergence, and allowing higher learning rates. Activate ReLU, introduce necessary nonlinearity, and support complex function modeling through activation based on positive input values. In order to manage high dimensions and optimize computational efficiency, a 2×2×2 maximum pooling layer is inserted after the second and fourth convolution layers to highlight the key features of classification while reducing the amount of data. The model has a fully connected layer and an output layer, with an initial learning rate of 0.01, and is designed to run up to 40 epochs. At the level of a single bacterium, according to the 3D-CNN architecture, the accuracy of the 3D hyperspectral method is 91.7%, while the input accuracy of 2D raw data is only 79.3%. This suggests that the 3D hyperspectral method has great advantages in the accuracy of microbial diagnosis.

[0035] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A single bacteria spectral detection method based on a filter array, characterized in that: The following steps are involved: Design and use a 5 × 5 filter array covering the spectral range of 614 to 974 nm to capture hyperspectral images of bacteria; During the detection process, a microscope objective with a magnification of 60 times and a numerical aperture of 1.25 was used to capture microscopic images of bacteria to obtain high-resolution bacterial imaging; The bacterial samples are guided to flow through the microfluidic channel in sequence, and a hyperspectral camera based on the filter array is used to continuously image the bacterial samples at a fixed position to quickly capture the dynamic changes of the bacteria; The captured bacterial mosaic images are reconstructed into three-dimensional hyperspectral data through super-resolution demosaicing technology to enhance the spatial details of the images and reveal the key structural features of the bacteria; A 3D convolutional neural network model that integrates bacterial structure and spectral features is used to analyze the three-dimensional hyperspectral data to achieve efficient and accurate detection at the single bacteria level.

2. The method according to claim 1, characterized in that The super-resolution demosaicing technique effectively enhances the spatial details of bacteria by reconstructing sparse data and amplifying the spectral signal by 8 times, revealing the key structural features of bacteria. It is achieved through weighted convolution, in which each missing pixel value is filled by averaging the adjacent known pixel values.

3. The method according to claim 2, characterized in that The technology uses bilinear interpolation technology to estimate the missing pixel values ​​and create a preliminary image that emphasizes low-frequency details. Specifically, for the target pixel, its four adjacent pixels in the original image are first determined; then, based on the position of the target pixel relative to these four known pixels, a specific weight calculation formula is used to estimate the value of the target pixel.

4. The method according to claim 1, characterized in that: A 3D CNN model is constructed to make up for the loss of high-frequency details that may be caused by bilinear interpolation. It can efficiently process hyperspectral data cubes to enhance the high-frequency details of images and improve the accuracy of detection.

5. The method according to claim 1, characterized in that The loss function is the weighted sum of the mean square error MSE and the structural similarity index SSIM, with weights w1=0.8 and w2=0.2 respectively. By optimizing this loss function, the model can reduce pixel-level errors and preserve the structural integrity of the image, achieving more accurate and visually consistent reconstruction.

6. The method according to claim 1, characterized in that The 3D convolutional neural network model can capture additional dimensions of spectral information, providing multiple spectral bands for each pixel, thereby increasing the data dimension while improving the accuracy of bacterial identification. The model starts with a series of convolutional layers configured with 3×3×3 convolutional kernels. While capturing detailed features, the model ensures that the resolution of the input data is fully preserved, avoiding information loss and extracting richer and more detailed features, providing strong feature support for subsequent bacterial classification and identification.

Images