Adaptive fluorescence signal reading method based on convolutional neural network

Abstract

The invention belongs to the technical field of information storage, and particularly provides a self-adaptive fluorescence signal reading method based on a convolutional neural network, which comprises the following steps: S1, hybridizing a DNA decoding chain with a coding chain in a micro pool of a storage substrate to obtain a primary image group of the hybridized storage substrate, and marking a fluorescence probe at the tail end of the decoding chain; s2, segmenting the primary image according to a micro-pool unit, extracting a plurality of sub-images, and performing data enhancement to obtain a secondary image; s3, marking the secondary image; s4, dividing the data set into a training set and a verification set; s5, inputting a training set into the CNN model for training, and assisting dynamic adjustment and optimization of hyper-parameters; and S6, applying the trained CNN model to test a secondary image, and outputting a fluorescence recognition result. According to the method, the judgment standard can be automatically learned according to a large number of training samples, the fluorescence signal can be accurately, stably and efficiently read, manual threshold value presetting is not needed, the influence of interference factors is small, and the robustness and accuracy of DNA information storage can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information storage, and particularly relates to a self-adaptive fluorescent signal reading method based on a convolutional neural network. BACKGROUND

[0002] Traditional DNA storage information reading adopts synthesis and sequencing.

[0003] In the DNA short-chain hybrid information storage method, the information reading mode adopts an image recognition algorithm. The mode is to obtain the luminescence image of the fluorescent probe in the micro-pool of the information storage substrate under laser irradiation, analyze the color and brightness intensity of the image by using image analysis software, and then determine whether there is a corresponding fluorescent probe in each micro-pool according to the color and brightness intensity. In this process, the accuracy, reliability and effectiveness of information reading are affected by various interference factors such as complex background, light fluctuation and color shift, and recognition errors are prone to occur, resulting in a decrease in accuracy and stability. SUMMARY

[0004] The image analysis method of the prior art is affected by interference factors such as complex background, light fluctuation and color shift, and has insufficient accuracy, reliability and effectiveness, and low efficiency.

[0005] To solve the above technical problems, according to some embodiments, the present application provides a self-adaptive fluorescent signal reading method based on a convolutional neural network, comprising: S1, hybridizing a DNA decoding chain with an encoding chain in a micro-pool of a storage substrate to obtain a primary image group of the hybridized storage substrate, wherein the decoding chain is labeled with a fluorescent probe at the end; S2, processing the primary image, segmenting the primary image according to the micro-pool and extracting a plurality of sub-images, and performing data enhancement on all the sub-images to obtain secondary images; S3, labeling each secondary image to generate a fluorescent label and construct a data set; S4, dividing the data set into a training set and a validation set according to a proportion; S5, constructing a CNN model, inputting the training set into the CNN model for multiple rounds of iterative training, and using the validation set to evaluate the performance of the CNN model in the training process to assist in dynamically adjusting and optimizing the hyperparameters; S6, applying the trained CNN model to test the secondary images, performing feature recognition on each micro-pool secondary image, and outputting a fluorescent recognition result, thereby automatically reading the fluorescent signal of each micro-pool.

[0006] Further, the acquisition process of the primary image group comprises: A storage substrate with data storage micropools containing encoding chains is prepared, the encoding chains have multiple types; the decoding chains have multiple types, the same type of decoding chain is marked by the same type of fluorescent probe, and the different decoding chains are marked by different colors of fluorescent probes; wherein each decoding chain has a unique type of encoding chain that is complementary to it; A decoding chain is added to the storage substrate for hybridization; The unhybridized decoding chain in the micropool is washed away with clean water; The storage substrate is irradiated with laser corresponding to the wavelength of the fluorescent probe, and a primary image is obtained; The storage substrate is washed with a washing solution to elute the hybridized decoding chain in the micropool, so as to remove the trace of the fluorescent probe; The above hybridization, washing, photographing and washing steps are repeated in turn for the remaining types of decoding chains to obtain multiple primary images corresponding to multiple decoding chains.

[0007] Further, the acquisition process of the primary image set comprises: A storage substrate with data storage micropools containing encoding chains is prepared, the encoding chains have multiple types; the decoding chains have multiple types, the same type of decoding chain is marked by the same type of fluorescent probe, and the different decoding chains are marked by different colors of fluorescent probes; wherein each decoding chain has a unique type of encoding chain that is complementary to it; Multiple decoding chains are added to the storage substrate for hybridization; the unhybridized decoding chain in the micropool is washed away with clean water; The storage substrate is irradiated with laser corresponding to the wavelength of the fluorescent probe, and a primary image is obtained;

[0008] Further, the label includes fluorescent probe sequence number, brightness value and presence or absence of fluorescence; The label of the fluorescent probe sequence number represents the label of the fluorescent probe type in the secondary image; The label of the brightness value represents the fluorescence intensity value of the micropool corresponding to the secondary image.

[0009] Further, after generating the label, the label is encoded, including: The fluorescent probe type in the fluorescent probe sequence number label is mapped to a number to obtain a digital label, and the digital label is converted into a multi-dimensional vector; The brightness value label is set as a continuous numerical value label, and the brightness value label is standardized during training to improve training stability.

[0010] Further, the primary image is segmented according to the micropool and multiple sub-images are extracted, including: segmenting the primary image according to the micropool, extracting each micropool as a 64×64 pixel picture, and converting the RGB value of each micropool pixel. Data augmentation is performed on all sub-images, including rotation, translation, scaling, and mirror flipping operations.

[0011] Further, the CNN model comprises an input layer, a first convolutional module, a second convolutional module, a Dropout layer, a fully connected layer, and an output layer connected in sequence. The first convolutional module comprises a first convolutional layer, an activation layer, and a max-pooling layer; the first convolutional layer is provided with 32 3x3 convolutional kernels, a stride of 1, and uses a ReLU activation function, followed by a 2x2 max-pooling layer. The second convolutional module comprises a second convolutional layer, an activation layer, and a max-pooling layer; the second convolutional layer is provided with 64 3x3 convolutional kernels, a stride of 1, and uses a ReLU activation function, followed by a 2x2 max-pooling layer.

[0012] The Dropout layer is used to prevent overfitting.

[0013] Further, the fluorescent recognition result comprises the fluorescent probe serial number, the brightness value, and the presence or absence of fluorescence in the micro-pool.

[0014] Further, the multi-round iterative training comprises: In the fluorescent probe serial number recognition process, a Softmax activation function is used, and cross-entropy is used as the loss function, so that the output value is a multi-dimensional vector corresponding to the fluorescent probe serial number. In the brightness value recognition process, a linear activation function is used to calculate the micro-pool brightness value, and an output is obtained as a real number.

[0015] Further, the multi-round iterative training further comprises: In the recognition process of the presence or absence of fluorescence: the judgment result of the presence or absence of fluorescence is taken as a learned classification task, and a softmax activation function is used to make the CNN model judge whether the fluorescent probe is present or absent through the training set. Or, in the recognition process of the presence or absence of fluorescence: calculate the total mean M1 of the brightness values of all training sets, and the mean M2 of the brightness values of strong signals and the mean M3 of the brightness values of weak signals in the training set; wherein the strong signals are the training sets with brightness values greater than the total mean in the training set, and the weak signals are the training sets with brightness values less than or equal to the total mean in the training set. Calculate the dynamic threshold T, T = a * M2 + b * M3, where a and b are adaptive weights, and a + b = 1. According to the variance σ of the training set total And the separation index Optimize the weights a and b, and the optimization process includes: traversing a in the range of [0.2, 0.4] with a step size of 0.001, calculating the corresponding S(a) value, and selecting a = arg max S(a); If the brightness value of the i-th micro-pool in the training set is greater than the dynamic threshold, it is judged as having fluorescence; if the brightness value of the i-th micro-pool in the training set is less than or equal to the dynamic threshold, it is judged as not having fluorescence.

[0016] The above technical solutions of the present application have at least the following beneficial technical effects: The adaptive fluorescence signal reading method based on the convolutional neural network in the present application first performs image segmentation, normalization, data enhancement and other processing on the primary image to obtain secondary images in units of micro-pools; then labels the secondary images with the fluorescence probe serial number, brightness value and label of presence or absence of fluorescence, and constructs a data set; the data set is used to perform multi-round iterative training on the CNN model, and the completed CNN model is used to test the secondary images, feature recognition is performed on each micro-pool secondary image, and the judgment results of the fluorescence probe serial number, brightness value and presence or absence of fluorescence are outputted, and the reading of the fluorescence signal of each micro-pool is completed. The model trained by the method can automatically learn the judgment standard according to a large number of training samples, can accurately, stably and efficiently read the fluorescence signal, does not need manual preset threshold, is less affected by interference factors, and can effectively improve the robustness and accuracy of DNA information storage. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any inventive labor.

[0018] Figure 1 is a flowchart of an adaptive fluorescence signal reading method based on a convolutional neural network in an embodiment of the present application.

[0019] Figure 2 is a correspondence relationship diagram of the fluorescence signal color of the micro-pool and the binary information in an embodiment of the present application.

[0020] Figure 3 is a CNN model structure diagram in an embodiment of the present application.

[0021] Figure 4 is a convolution calculation process diagram in an embodiment of the present application.

[0022] Figure 5 is a direct effect diagram drawn according to the detection of the fluorescence probe serial number and the brightness value in an embodiment of the present application.

[0023] Figure 6 is a direct effect diagram drawn according to the judgment result of detecting the sequence number, brightness value and presence or absence of fluorescence of the fluorescent probe in one embodiment of the present application. DETAILED DESCRIPTION

[0024] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the drawings. However, those skilled in the art can understand that in the embodiments of the present application, many technical details are proposed in order to make the readers better understand the present application. However, the technical solutions claimed by the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the following embodiments is for the convenience of description and should not constitute any limitation on the specific implementation of the present application. The embodiments can be combined with each other and mutually quoted without contradiction.

[0025] First, the basic principle of DNA short chain hybridization is explained: the DNA storage substrate is provided with a plurality of micro-pools, and the micro-pools store DNA coding chain combinations. To obtain the information stored by the DNA storage substrate, the information of the DNA coding chain in each micro-pool needs to be obtained first, and then the information of the DNA coding chain in all micro-pools is combined in order to form the complete stored information. The DNA coding chain in the micro-pool is a single strand, which can be a combination of one or more coding chains, or zero.

[0026] In order to understand which coding chains are in the micro-pool, a decoding chain that is uniquely complementary to the possible coding chains needs to be added to the micro-pool. There are many decoding chains, and each decoding chain carries a fluorescent probe. Add one decoding chain (hundreds of millions) to the micro-pool. If the micro-pool contains a complementary coding chain, a stable DNA double strand will be formed by the reaction. At this time, the micro-pool is washed with water, and the decoding chain that has not reacted can be washed away. The stable DNA double strand obtained by the reaction of the decoding chain and the coding chain will not be washed away by water. After the double strand is formed, the fluorescent probe carried by the decoding chain can emit fluorescence under the irradiation of laser light of the corresponding wavelength. According to the fluorescence in the micro-pool, it can be judged whether there is a decoding chain in the micro-pool, and then the type of complementary coding chain in the micro-pool is obtained. According to the combination of the coding chain types in the micro-pool, the original storage information can be restored. How to express the storage information will be explained by examples below.

[0027] As shown in the flowchart of Figure 1 , one embodiment of the present application provides a self-adaptive fluorescent signal reading method based on a convolutional neural network, comprising: S1, hybridize the decoding strand of DNA with the encoding strand in the micro-pool of the storage substrate, obtain a primary image set of the hybridized storage substrate, and the micro-pool to be judged in the primary image set has characteristics such as fluorescent color and brightness; the decoding strand is labeled with a fluorescent probe, the same kind of decoding strand is labeled with the same kind of fluorescent probe, and the fluorescent probes of different decoding strands are different in color; wherein, there is only one kind of encoding strand complementary to each decoding strand. It should be noted that the characteristics of the primary image set are not judged in this step, and the hybridization is between the decoding strand (carrying the fluorescent probe) and the encoding strand. Hybridization is an operation step, not a reaction. Whether the reaction occurs is determined by whether there is a complementary encoding strand in the micro-pool.

[0028] S2, segment the primary image by micro-pool and extract a plurality of sub-images, perform data enhancement on all sub-images to obtain secondary images; through these processes, the robustness and generalization ability of subsequent model training can be improved.

[0029] S3, label each secondary image to generate a fluorescent label and construct a data set.

[0030] S4, divide the data set into a training set and a validation set in proportion, wherein the training set is used to train a CNN model (convolutional neural network model), and the validation set is used to evaluate the performance of the model during training and assist in dynamically adjusting and optimizing the hyperparameters.

[0031] S5, construct a CNN model, input the training set into the CNN model for multiple rounds of iterative training, and use the validation set to evaluate the performance of the CNN model during training and assist in dynamically adjusting and optimizing the hyperparameters.

[0032] S6, apply the trained CNN model to test the secondary image, perform feature recognition on each micro-pool secondary image, and output the fluorescent recognition result, thereby completing the automatic reading of the fluorescent signal of each micro-pool.

[0033] The adaptive fluorescent signal reading method based on the convolutional neural network of the present application hybridizes the decoding strand of DNA with the encoding strand in the micro-pool of the storage substrate, obtains a primary image set of the hybridized storage substrate, and then performs segmentation, data enhancement, etc. on the primary image to obtain secondary images with micro-pools as units; then label the secondary images with fluorescent labels to construct a data set; use the data set to train the CNN model for multiple rounds of iteration, use the trained CNN model to test the secondary image, perform feature recognition on each micro-pool secondary image, output the fluorescent recognition result, and complete the reading of the fluorescent signal. This method can automatically learn the judgment standard according to a large number of training samples, accurately, stably and efficiently read the fluorescent signal, does not need manual preset threshold, is less affected by interference factors, and can effectively improve the robustness and accuracy of DNA information storage.

[0034] The following specific embodiments are further illustrations of this application. The examples given do not represent all the implementation methods of this application; only some are used as examples for illustration. Specific embodiments are as follows: In step S1, there are two hybridization methods: S11, multiple decoding chains are hybridized with the coding chains in the micro pool in batches; S12, multiple decoding chains are simultaneously mixed and hybridized with the coding chains in the micro pool.

[0035] The first hybridization method, S11, involves hybridizing multiple decoding strands with the coding strand of the micropool in batches. Taking an example with four types of coding strands, a primary image set of the storage micropool after hybridization of the DNA decoding strand and the coding strand is obtained, including: S111. Prepare a substrate with a 96-well storage micropool array, where each well is a micropool that can store short DNA strands, and each micropool can be written with four designed DNA coding strands.

[0036] S112. Prepare four decoding strands that are uniquely complementary to the DNA coding strand. The ends of the different decoding strands are labeled with different fluorescent probes. The fluorescent probes labeled with the same decoding strand have the same color, while the fluorescent probes labeled with different decoding strands have different colors.

[0037] S113. Add the first decoding chain to each micropool for hybridization.

[0038] S114. After a certain period of time (the time required for hybridization to complete), wash the microcell multiple times with clean water to remove the unhybridized decoding strands and fluorescent probes. (It is possible that some or all of the decoding strands have reacted, forming stable double strands, and will not be washed away during the water washing process. The fluorescent probes carried by the decoding strands will also not be washed away, indicating that there is a complementary coding strand in the microcell. Alternatively, it is possible that none of them have reacted, indicating that there is no complementary coding strand in the microcell, and no stable double strands have been formed. In this case, the decoding strands and their fluorescent probes will be completely washed away with clean water. However, the coding strands will not be washed away regardless of whether they have reacted or not.)

[0039] S115. The storage substrate is irradiated with a laser corresponding to the wavelength of the fluorescent probe to induce fluorescence in the microcells. A fluorescence microscope is then used to capture a primary image corresponding to the fluorescent probe carried by the first decoding strand. If the DNA coding strand of the microcell reacts and binds to the decoding strand carrying the probe, forming a double strand, the reacting decoding strand and fluorescent probe will not be washed away within the double strand, and the fluorescent probe will be present in the primary image. If no coding strand in the microcell matches the first decoding strand, and no stable double strand is formed, the fluorescent probe will be washed away with water along with the decoding strand, and the primary image will not show any fluorescence signal from the fluorescent probe. It should be noted that even without a reaction, the storage substrate itself will reflect the laser and will have a certain brightness value.

[0040] S116, after taking a photo, wash the hybridized micro-pool multiple times to elute the hybridized decoding chain in the micro-pool. The washing liquid can be formamide. This step is different from the previous water washing process. The purpose of washing is to separate the DNA double strands formed in the previous step, remove the hybridized decoding chain, remove the traces of fluorescent probes, and keep the encoding chain in the micro-pool, so as to restore the initial state of the micro-pool for the next hybridization or information reading.

[0041] S117, sequentially add the remaining decoding chains to the storage substrate (each micro-pool) for hybridization, cleaning, photographing, and washing to obtain a primary image corresponding to each decoding chain.

[0042] In order to meet the data requirements of the subsequent CNN model, multiple primary image groups can be prepared under different environmental conditions and equipment conditions.

[0043] The second hybridization method: S12, multiple decoding chains are mixed and hybridized with the encoding chains in the micro-pool to prepare for reading the data on the storage substrate.

[0044] S121, add multiple decoding chains to the storage substrate for hybridization with the encoding chains in the micro-pool; the decoding chain is labeled with a fluorescent probe, and the fluorescent probes labeled by different decoding chains have different colors. Each decoding chain has a unique encoding chain that is complementary to it.

[0045] S122, wash away the unhybridized decoding chains in the micro-pool with water.

[0046] S123, use a laser corresponding to the wavelength of the fluorescent probe to irradiate the storage substrate in sequence, and take a photo to obtain a primary image corresponding to each decoding chain.

[0047] After taking a photo, wash it again to separate the double strands formed by hybridization, wash away the decoding chain, and restore the initial state of the micro-pool for the next reading.

[0048] The first hybridization method is to wash, take a photo, and wash after each decoding chain is hybridized. One fluorescent probe hardly affects the primary image of another fluorescent probe. The operation steps are more, and the accuracy of each step needs to be guaranteed, but the final reading accuracy is high and the required data set is small. The second hybridization method is to add multiple decoding chains to the micro-pool of the storage substrate at the same time, let the decoding chains hybridize with the encoding chains, then wash away the unhybridized decoding chains, then select the laser wavelength corresponding to the decoding chain probe in sequence, and take a photo in sequence. This method has a certain degree of mutual influence among multiple fluorescent probes, requires a large amount of data set, but has fewer operation steps, which can effectively shorten the information reading time.

[0049] One of the purposes of the present application is to provide an algorithm to accurately determine the number of fluorescent probes in each micro-pool, to determine which decoding chains are successfully hybridized, and to determine the combination of encoding chains in the micro-pool. Reading the fluorescent signal feedback of the fluorescent probe is one of the effective ways to determine the number of fluorescent probes. Here, the use of reading the fluorescent probe in the micro-pool (expression storage) is described: taking the hybridization of four decoding chains and four encoding chains as an example, in the S1 step, four decoding chain hybridization tests are performed on each micro-pool. The hybridization methods include mixed hybridization of four decoding chains and encoding chains in the micro-pool, and batch hybridization of four decoding chains and encoding chains in the micro-pool. Taking the mixed hybridization of four decoding chains and encoding chains in the micro-pool as an example, in the obtained primary image, there is one fluorescent probe, or multiple fluorescent probes, or no fluorescent probe in the micro-pool. According to the order of the decoding chain corresponding to the fluorescent probe, if there is a certain fluorescent probe in the micro-pool, it indicates that there is an encoding chain corresponding to the decoding chain, which is represented by "1", and if there is no corresponding fluorescent probe, it is represented by "0". For example, Figure 2 As shown in the four primary images, the four fluorescent probes exist in the first micro-pool, so the binary number represented by the micro-pool is "1111"; the second micro-pool has the second and fourth fluorescent probes, so it represents "1010"; and the third micro-pool has only the first fluorescent probe, so it represents "0001". In this way, each micro-pool represents a four-bit binary information. According to the ordering rules of all micro-pools on the substrate, these binary numbers can be combined to express more ordered storage information.

[0050] During the acquisition of the primary image, the substrate body, residues, the concentration of the washing solution, and environmental light interference all affect the reading of the fluorescent signal. Therefore, the present application improves the accuracy and robustness of reading the fluorescent probe through subsequent processing, model training, and model processing steps.

[0051] S2, processing each primary image: processing the primary image, automatically segmenting and extracting multiple sub-images from the primary image according to the micro-pool unit, specifically: S21, dividing each primary image into multiple sub-images in the form of a micro-pool unit.

[0052] S22, adjusting the multiple sub-images to a size of 64x64, and converting the RGB values of each sub-image micro-pool pixel, so that the brightness of different sub-images is compared under the same standard. The maximum and minimum value standard normalization method is used to scale the data of each pixel value in the secondary image, which conforms to the following formula:

[0053] wherein, X represents the original pixel value recognized by the algorithm, X min is the minimum pixel value of the image, 0,X max 255 for the maximum pixel value of the image.

[0054] S23, image enhancement includes: random rotation, translation, scaling and lens flipping operation are performed on all sub-images to increase the diversity of data, which can improve the robustness of the CNN model in the subsequent training process.

[0055] S3, label each secondary image to generate a fluorescence label and build a dataset. The basis for labeling is that the wavelength of the laser irradiating the micro-pool is known, and the corresponding fluorescence color is also known. The existing graphic analysis software can obtain the brightness value of each micro-pool to confirm whether there is corresponding fluorescence in the micro-pool. The label specifically includes: S31, each secondary image is labeled with a fluorescent probe serial number, a brightness value, and a label indicating whether there is fluorescence.

[0056] The fluorescent probe serial number label represents the label of the fluorescent probe species of the secondary image. Alternatively, the fluorescent probe serial number can also represent the fluorescence color of different probes.

[0057] The brightness value label represents the fluorescence intensity value of the secondary image.

[0058] The label indicating whether there is fluorescence represents whether there is fluorescence or no fluorescence in the micro-pool of the secondary image.

[0059] S32, label encoding: convert the above labels into numerical labels that can be understood by the neural network. Since the convolutional neural network (CNN) is a classification model, the information of different fluorescent probe serial numbers and brightness values is digitized through label encoding processing to provide effective input data for the CNN model. Specifically as follows: Fluorescent probe serial number label encoding: the label is a label for the secondary image. Taking four kinds of fluorescent probes as an example, the serial numbers of the four fluorescent probes are set to 1 (First), 2 (Second), 3 (Third), and 4 (Fourth), respectively. The case of no fluorescent probe is also included, and the serial number of the no fluorescent probe is set to 0 (None), becoming a digital label, and then each digital label is converted into a five-dimensional vector using one-hot encoding. The fluorescent probe serial number label is mapped to the following numerical value: color_map = {"First": 0, "Second": 1, "None": 2, "Third": 3, "Fourth": 4}.

[0060] Brightness value label encoding: keep it as a continuous numerical value representing the intensity of fluorescence, with a value between 0 and 255.

[0061] S4, divide the dataset into training set and validation set in proportion, wherein the training set accounts for 80%, and the validation set accounts for 20%.

[0062] S5, build a CNN model, input the training set into the CNN model for multi-round iterative training, evaluate the performance of the model in the training process by using the validation set, and assist in completing the dynamic adjustment and optimization of the hyperparameters, including the number of layers of the neural network, the number of neurons in each layer, the activation function and other network architecture related parameters, the learning rate, the batch size of training, the optimization algorithm, the optimizer parameters, the number of iterations and other parameters related to the optimizer.

[0063] S51, the structure of the built CNN model follows the classic hierarchical structure of convolutional neural network, and through the combination of each layer such as convolutional layer, the network can efficiently extract image features, and classify and regress according to the extracted features.

[0064] As shown in Figure 3 , the CNN model includes an input layer, a first convolutional module, a second convolutional module, a Dropout layer, a fully connected layer and an output layer connected in sequence.

[0065] The input layer is used to receive the dataset and perform subsequent processing.

[0066] The input layer is connected with two convolutional modules (the first convolutional module and the second convolutional module), and each convolutional module is connected with an activation layer and a pooling layer after the convolutional layer. The role of the convolutional layer is to extract low-level features of the image such as edges and textures through the convolution kernel; the activation layer is closely connected with the convolutional layer, and increases the nonlinear ability of the model through the activation function; the pooling layer reduces the size of the feature map through the maximum pooling operation, reduces the complexity of the calculation, and enhances the robustness of the model. Finally, the CNN model outputs the recognition results of the fluorescence probe serial number, the brightness value and the presence or absence of fluorescence through the output layer.

[0067] After inputting the data set, the convolutional layer can extract low-level features from the secondary image. Among them, the first convolutional module includes a convolutional layer, an activation layer and a maximum pooling layer; the convolutional layer in the first convolutional module uses 32 3x3 convolutional kernels, the stride is 1, and the activation function is ReLU. After the image data passes through the first convolutional layer, the model is processed by the maximum pooling layer with a 2x2 pooling window for dimension reduction. The second convolutional module includes a second convolutional layer, an activation layer and a maximum pooling layer; the convolutional layer in the second convolutional module uses 64 3x3 convolutional kernels, the stride is also 1, and the activation function ReLU further extracts deeper features, such as middle-level semantic features such as the complete shape of fluorescent spots, color space structure, local contrast mode and noise features. Then the feature compression is performed through the maximum pooling layer. The design and selection of the convolutional kernel size and the pooling layer aims to balance the calculation efficiency and the feature extraction ability, so that the model can effectively learn meaningful features under a smaller image size. The convolution calculation formula is as follows:

[0068] wherein, W is the convolution kernel, x is the input data, b is the bias term. The convolution calculation process is shown in Figure 4 .

[0069] The activation function ReLU calculation formula is as follows:

[0070] This function sets negative values to zero and keeps positive values unchanged, thereby improving the gradient propagation efficiency.

[0071] The maximum pooling calculation formula of the 2x2 pooling window is as follows:

[0072] That is, the maximum pixel value in the 2x2 pooling window is selected as the output of the region, and the feature extraction is completed.

[0073] A Flatten layer is provided between the convolutional layer and the fully connected layer. After the feature extraction stage is completed, the model uses the Flatten layer to flatten the multi-dimensional feature map output by the convolutional layer into one-dimensional data to provide input for the fully connected layer. At the same time, a Dropout layer is added before the fully connected layer, which randomly discards a certain proportion of neurons to prevent overfitting, thereby reducing the model's excessive dependence on training data. The fully connected layer connects all neurons to all outputs of the previous layer, which helps the model to map image features to the final classification and regression results.

[0074] The output layer is composed of three parts, which respectively output the recognition results of the fluorescent probe serial number, the brightness value and the presence or absence of fluorescence, and the code is as follows: output_color = Dense(5, activation='softmax', name='color')(x) output_brightness = Dense(1, activation='linear', name='brightness')(x) output_judgment = Dense(1, activation='sigmoid', name='judgment')(x).

[0075] S52. Fluorescent probe number prediction is a multi-class classification task. It uses the Softmax activation function, resulting in a five-dimensional vector as the output. Each vector represents the probability of five possible outcomes for detecting four probes within a storage micropool. For each secondary image, the CNN extracts features step-by-step through convolutional layers, then outputs the probabilities of the five outcomes. Finally, the probe with the highest probability is selected as the predicted fluorescent probe number.

[0076] The formula for the Softmax activation function is as follows: .

[0077] S53. Brightness value prediction is a continuous value regression task. This application uses a linear activation function so that the output value is a real number representing the brightness value of the secondary image. During training, the model adjusts the weights by minimizing the error between the predicted brightness and the actual brightness, thereby maximizing prediction accuracy. After training, the CNN can predict brightness based on the features it has learned. The specific steps are as follows: a. Forward propagation: After the input secondary image is processed through convolutional layers, activation layers, pooling layers, and fully connected layers, a real value is finally output in the output layer using a linear activation function, representing the predicted brightness.

[0078] b. Since brightness is a continuous value, the mean square error loss function is used. L :

[0079] It is the first i The predicted brightness value of an image. y i This is the actual brightness value. n It represents the number of samples; the smaller the loss value, the more accurate the model's prediction.

[0080] c. Backpropagation: Through the backpropagation algorithm, the model automatically calculates the loss function for all model parameters (such as convolutional kernel weights).w , bias b ) gradient:

[0081] d.Adam optimizer: makes the model can adjust the parameters along the "let the loss decrease the fastest direction", so as to approach the correct prediction value faster and more stably.

[0082] e. Repeat until convergence: the above process is repeated for several rounds until the model's predicted brightness value is very close to the true value, and the loss tends to be minimal.

[0083] S53, the prediction of whether there is fluorescence; the following two implementation schemes can be taken.

[0084] The first scheme: The CNN model takes "judgment" as a learned classification task, and through the use of softmax activation function, the model automatically learns to judge whether the fluorescent probe exists through the training set and the corresponding fluorescent probe sequence number label, thereby saving human intervention.

[0085] The second scheme: S531, calculate the total mean M1 of the brightness values of all training sets, and the mean M2 of the brightness values of strong signals and the mean M3 of the brightness values of weak signals in the training set; wherein the strong signal is the training set with brightness value greater than the total mean in the training set, and the weak signal is the training set with brightness value less than or equal to the total mean in the training set.

[0086] The total mean M1 meets the formula: , wherein x i is the brightness value of the fluorescence signal of the micro-pool of the ith training set, and n is the total number of training sets; The mean M2 of the brightness values of strong signals meets the formula: .

[0087] The mean M3 of the brightness values of weak signals meets the formula: .

[0088] S532, calculate the dynamic threshold T, T = a* M2 + b* M3, wherein a and b are adaptive weights, and a + b = 1; S533, according to the variance σ total and the separation index of the training set, optimize the weights a and b, wherein the calculation formula of the variance of the training set is σ total =Sqr(∑(x iM2 / (n-1)), Sqr represents square root; the optimization process includes: traversing a in the range of [0.2, 0.4] with a step of 0.001, calculating the corresponding S(a) value, selecting a = arg max S(a), that is, selecting a that makes S(a) maximum as the adaptive weight.

[0089] In S534, if the brightness value of the training set is greater than the dynamic threshold T, it is judged as having fluorescence; if the brightness value of the training set is less than or equal to the dynamic threshold T, it is judged as not having fluorescence.

[0090] Optionally, in step S533, the values of a and b can also be directly set: calculating the dynamic threshold T, T = a * M2 + b * M3, wherein a and b are preset weights, satisfying a + b = 1, and 0.2 ≤ a ≤ 0.4.

[0091] Through tests, the judgment result of the first scheme is more ideal and accurate in the case of sufficient training data, and the judgment of the second scheme is relatively accurate in the case of less training data.

[0092] The model divides the data set into a training set and a validation set using the train_test_split function, that is, 80% of the data in the data set is used for training, and 20% of the data is used for verifying the training effect. After division, the data of the training set and the validation set are stored in X_train, X_val and the related parts of the labels (y_color_train, y_color_val, brightness_train, etc.). This data division can effectively avoid model overfitting and provide reliable evaluation results. The training process is performed by calling the model.fit function, the model learns on the training set and evaluates on the validation set, and the training progress is monitored in real time. Through learning on the data set, the model can automatically identify the fluorescence probe serial number, the brightness value, and judge whether there is fluorescence in the secondary image.

[0093] Model optimization: the CNN model also uses the cross-entropy loss function (Cross-Entropy Loss) for classification tasks, which can measure the difference between the probe category distribution output by the model and the label. In the fluorescence probe serial number prediction classification task, the cross-entropy loss function can effectively guide the model optimization, ensuring that the probability of each output category is as close to the true label as possible. Its loss formula is:

[0094] wherein, y i is the true label, The predicted probabilities are the results outputted by the Softmax activation function. The goal of the cross-entropy loss function is to minimize the difference between the predicted probabilities and the true labels, ensuring that the model accurately predicts the probability distribution for each class. When predicting the secondary image (which can not be the secondary image in the dataset), using the cross-entropy loss function can obtain the probability of the secondary image having the probe sequence number of the predetermined five cases based on the experience gained during the training process of the dataset, and decide that the case with the maximum probability is the probe sequence number of the secondary image. By optimizing the cross-entropy loss, the model can continuously adjust the weights to output probabilities closer to the true labels for a given input.

[0095] During the training process, the gradients are calculated using backpropagation, and the model parameters are updated using the Adam optimizer, which combines gradient descent and momentum and uses historical gradient information to dynamically adjust the learning rate of each parameter. The core is to estimate the first moment (average gradient) and second moment (square of gradient) of the gradient to dynamically adjust the learning rate. The update rule is as follows: (1) Initialize parameters: initialize the parameters of the convolutional model, including the convolution kernel weights w and the bias b .

[0096] (2) Calculate the gradient: in each iteration, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm. For the convolution kernel weights w and the bias b , their gradients are calculated respectively.

[0097] (3) Update the first moment estimate and the second moment estimate: for each parameter w or b , at the t th iteration, the first moment estimate m t and the second moment estimate v t are updated as follows: ,

[0098] where m t is the first-order estimate of the gradient, i.e., the average of the gradient, v t is the second-order moment estimate, i.e., the first-order estimate of the square of the gradient; β 1 and β 2 are decay coefficients, usually β 1 set to 0.9, β 2Set to 0.999, so that the dependence on history gradients is longer and the variance of the updates is slightly decreased; θ Represent w Or b .

[0099] (4) Since the initial time m t And v t The bias is present in the early estimates, and to correct this bias, Adam will correct the bias of the first and second moments by the following formula: ,

[0100] (5) Use the corrected moment estimate to update the model parameters, update the network parameters based on the corrected estimate θ t , the formula is:

[0101] In the formula, α is the learning rate, ε is a small constant (can be 10 -8 ), a smoothing term to prevent division by zero, That is, the gradient of the loss function.

[0102] In the model training phase, use the fit method to train the model, and set the training period (epochs) and batch size (batch_size). During training, not only the training set is used to train the model, but also the validation set is used to evaluate the performance of the model.

[0103] To ensure that the training log during training can be recorded, use the CSVLogger callback function to save the loss value and accuracy information of each round of training to the CSV file. At the same time after the training is completed, use the saved model to test and verify the performance of the model on unknown data. Through the model.predict() method, the prediction results of the model on the secondary image can be obtained, including the fluorescence probe serial number, brightness value and judgment result of whether there is fluorescence.

[0104] When the CNN model is trained for a set period of time, input the secondary image for testing, and the conclusion based on the secondary image obtained by the CNN model after running analysis is as follows: Test Image Color: “ ”, Brightness: “ ”, Judgment: “ ” Wherein, the conclusion obtained by the CNN model analyzing the secondary image is the recognition result of a fluorescent probe sequence number, brightness value and presence or absence of fluorescence. Through the above method, high-precision reading of the fluorescence signal can be stably achieved, and the automation degree and accuracy of DNA data reading can be improved.

[0105] The application of the probe sequence number obtained from the micro-pool will be explained below. The CNN model outputs a corresponding fluorescent probe sequence number and a brightness value for each secondary image. The four secondary images corresponding to the same micro-pool (four kinds of decoding chain detection carrying fluorescent probes) can indicate which fluorescent probes and brightness values are present in the micro-pool. The fluorescent information output by each micro-pool is drawn into a visualized form similar to Figure 2 by manual or image software, and the fluorescence signals in multiple micro-pools on the well plate can be intuitively represented as Figure 5 . Figure 5 Only a small number of micro-pools in the above image do not detect fluorescence signals, and most micro-pools can detect corresponding fluorescent probe sequence numbers and brightness values.

[0106] In combination with the judgment result of the presence or absence of the fluorescent probe given by the CNN model, the reading result of the CNN model is shown in Figure 6 . It should be noted that Figure 4 , Figure 5 and Figure 6 are drawn according to the predicted data output by the CNN model, and are not output by the CNN model. When interpreting the storage information of the entire storage substrate, such a graph does not need to be drawn. Only the presence of which fluorescent probe sequence numbers in each micro-pool needs to be determined. The judgment of brightness value and presence or absence of fluorescence is an auxiliary judgment standard used by the CNN model to determine the fluorescent probe sequence number in order to overcome various interference factors such as complex background, light fluctuation and color shift, and to reduce errors.

[0107] In an embodiment of the present application, after the CNN model is trained, the primary image group is used to retrain the trained CNN model to predict the primary image group and output the fluorescent probe sequence number, brightness value and presence or absence of fluorescence in each micro-pool of the primary image.

[0108] The above is an embodiment of the reading method when four fluorescent probes (one-to-one corresponding to four decoding chains and four encoding chains) are used. The number of fluorescent probes (corresponding to decoding chains and encoding chains) can also be applied when the number of other types is used, such as 2 or 7 encoding chains in the micro-pool. Only the corresponding adjustment of the type and sequence number of the fluorescent probe label is required.

[0109] It should be understood that the foregoing detailed description of the application, rather than limiting the application, is intended to explain and describe the current application to those skilled in the art. Accordingly, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the application shall be included in the protection scope of the application. In addition, the claims of the present application are intended to cover all variations and modifications falling within the scope and boundary of the appended claims, or the equivalent form of such scope and boundary.

Claims

1. A convolutional neural network-based adaptive fluorescence signal reading method, characterized in that, The method comprises the following steps: S1, hybridize the decoding chain labeled with fluorescent probes to the encoding chain in the micro-pool of the storage substrate, obtain the primary image set of the hybridized storage substrate, and the decoding chain is labeled with fluorescent probes; S2, process the primary image, segment the primary image according to the micro-pool, extract a plurality of sub-images, and perform data enhancement on all sub-images to obtain secondary images; S3, label each secondary image to generate a fluorescent label and construct a data set; S4, divide the data set into a training set and a validation set according to a proportion; S5, construct a CNN model, input the training set into the CNN model for multiple rounds of iterative training, and use the validation set to evaluate the performance of the CNN model during the training process to assist in dynamically adjusting and optimizing the hyperparameters; S6, apply the trained CNN model to test the secondary images, perform feature recognition on the secondary images of each micro-pool, and output the fluorescent recognition result, thereby completing the automatic reading of the fluorescent signal of each micro-pool.

2. The method of claim 1, wherein, The acquisition process of the primary image set comprises: Prepare a storage substrate with storage micro-pools containing encoding chains, and there are multiple types of encoding chains; there are multiple types of decoding chains, the same type of decoding chain is labeled with the same type of fluorescent probe, and the fluorescent probes of different decoding chains are different in color; wherein each type of decoding chain has only one type of encoding chain that is complementary to it; Add a decoding chain to the storage substrate for hybridization; Rinse the decoding chain in the micro-pool that has not been hybridized with clean water; Irradiate the storage substrate with laser corresponding to the wavelength of the fluorescent probe and take a photo to obtain the primary image; Add a washing solution to clean the storage substrate and elute the hybridized decoding chain in the micro-pool to remove the traces of the fluorescent probe; Select the remaining types of decoding chains in turn and repeat the above hybridization, cleaning, photographing, and washing steps to obtain multiple primary images corresponding to multiple types of decoding chains.

3. The method of claim 1, wherein, The acquisition process of the primary image set comprises: Prepare a storage substrate with data storage micro-pools containing encoding chains, and there are multiple types of encoding chains; there are multiple types of decoding chains, the same type of decoding chain is labeled with the same type of fluorescent probe, and the fluorescent probes of different decoding chains are different in color; wherein each type of decoding chain has only one type of encoding chain that is complementary to it; Add multiple decoding chains to the storage substrate for hybridization; rinse the decoding chain in the micro-pool that has not been hybridized with clean water; Irradiate the storage substrate with laser corresponding to the wavelength of the fluorescent probe and take a photo to obtain the primary image corresponding to each type of decoding chain.

4. The method according to claim 2 or 3, characterized in that, The label includes the fluorescent probe sequence number, the brightness value, and the presence or absence of fluorescence; The fluorescent probe sequence number label indicates the labeling of the fluorescent probe type in the secondary image; The brightness value label indicates the fluorescence intensity value of the corresponding micro-pool of the secondary image.

5. The method of claim 4, wherein, After generating the label, encode the label, including: Map the fluorescent probe type in the fluorescent probe sequence number label to a number to obtain a digital label, and convert the digital label to a multi-dimensional vector; Set the brightness value label as a continuous numerical label, and standardize the brightness value label during training to improve training stability.

6. The method of claim 1, wherein The primary image is segmented by the micro-pool and a plurality of sub-images are extracted, including: segmenting the primary image by the micro-pool, extracting each micro-pool as a 64*64 pixel picture, and converting the RGB value of each micro-pool pixel; Data augmentation is performed on all sub-images, including: rotating, translating, scaling, and mirror flipping operations on all sub-images.

7. The method of claim 5, wherein, The CNN model comprises an input layer, a first convolution module, a second convolution module, a Dropout layer, a full connection layer, and an output layer connected in sequence; The first convolution module comprises a first convolution layer, an activation layer, and a maximum pooling layer; the first convolution layer is provided with 32 3*3 convolution kernels, a stride of 1, and uses a ReLU activation function, followed by a 2*2 maximum pooling layer; The second convolution module comprises a second convolution layer, an activation layer, and a maximum pooling layer; the second convolution layer is provided with 64 3*3 convolution kernels, a stride of 1, and uses a ReLU activation function, followed by a 2*2 maximum pooling layer; The Dropout layer is used to prevent overfitting.

8. The method of claim 7, wherein, The fluorescent recognition result comprises the fluorescent probe serial number, the brightness value, and the presence or absence of fluorescence in the micro-pool.

9. The method of claim 8, wherein, The multi-round iterative training comprises: In the fluorescent probe serial number recognition process, a Softmax activation function is used, and cross-entropy is used as a loss function, so that the output value is a multi-dimensional vector corresponding to the fluorescent probe serial number; In the brightness value recognition process, a linear activation function is used to calculate the micro-pool brightness value, and an output is obtained as a real number.

10. The method of claim 9, wherein, The multi-round iterative training further comprises: In the presence or absence of fluorescence recognition process: the presence or absence of fluorescence judgment result is used as a learned classification task, and a softmax activation function is used to make the CNN model judge whether the fluorescent probe is present or absent through the training set; Or, in the presence or absence of fluorescence recognition process: the total mean M1 of the brightness values of all training sets is calculated, as well as the mean M2 of the brightness values of strong signals in the training set and the mean M3 of the brightness values of weak signals in the training set; wherein the strong signals are the training sets with brightness values greater than the total mean in the training set, and the weak signals are the training sets with brightness values less than or equal to the total mean in the training set; A dynamic threshold T is calculated, T=a* M2+b* M3, wherein a and b are adaptive weights, and a+b=1; According to the variance σ of the training set total and the resolution index Optimizing the weights a and b, the optimization process includes: traversing a in the range of [0.2, 0.4] with a step size of 0.001, calculating the corresponding S(a) value, and selecting a = arg max S(a); If the brightness value of the i-th micro-pool in the training set is greater than the dynamic threshold, it is judged as having fluorescence; if the brightness value of the i-th micro-pool in the training set is less than or equal to the dynamic threshold, it is judged as having no fluorescence.

Images