Gene chip fluorescence signal analysis method and system based on deep learning

Abstract

The invention discloses a gene chip fluorescence signal analysis method and system based on deep learning, and belongs to the field of image extraction. The method comprises the following steps: acquiring a scanning image of the gene chip and preprocessing; constructing a multi-module deep learning model taking end-to-end prediction as a core, wherein the model can automatically learn multi-scale features of a microbead area in an image and perform signal analysis; according to the model, a multi-task learning framework is adopted, and microbead positioning, brightness classification and fluorescence signal intensity prediction are completed at the same time; automatically optimizing the feature extraction and signal analysis process through model training; and finally generating a fluorescence signal intensity value according to model output. According to the invention, automation and intellectualization of gene chip signal analysis are realized, different experiment conditions can be adapted, and the accuracy and robustness of analysis are obviously improved.

Description

Technical Field

[0001] This invention relates to image processing, and more particularly to a method and system for analyzing gene chip fluorescence signals based on deep learning. Background Technology

[0002] Gene chip technology, as an important high-throughput detection method in modern biological research, is widely used in gene expression analysis, mutation detection, and pathogen identification. Its core principle is to reflect the hybridization status of target molecules through fluorescence signal intensity. However, traditional fluorescence signal analysis methods have significant limitations. First, relying on manually set feature extraction rules, such as calculating the average pixel value in a fixed region, makes it difficult to adapt to changes in image characteristics under different experimental conditions. In practical applications, due to differences in probe density, hybridization efficiency, and scanning parameters, the fluorescence signal distribution exhibits complex spatial heterogeneity, and fixed rules can easily lead to insufficient feature extraction or amplified noise interference. Second, traditional methods typically employ a segmented processing flow, including independent steps such as bead localization, feature extraction, and signal calculation. The accumulation of errors at each stage reduces the overall accuracy of the analysis. Furthermore, manually designed feature weight allocation is highly subjective and lacks adaptability to the diversity of bead morphology and background noise, affecting the reliability and repeatability of the results.

[0003] In recent years, deep learning technology has made breakthroughs in image recognition, but its application in gene chip analysis still faces challenges. For example, general image processing models are not optimized for the spatial distribution characteristics of fluorescence signals, and most methods rely on large amounts of labeled data, while gene chip experiments are costly and have limited labeled samples. Furthermore, existing methods often focus on single tasks, such as classification or regression, failing to fully utilize the synergistic effects between multiple tasks. Therefore, there is an urgent need to develop a new method that can adaptively learn features, integrate multi-scale information, and optimize the end-to-end analysis process. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a gene chip fluorescence signal analysis method and system based on deep learning. By constructing an end-to-end deep learning model, it realizes intelligent analysis from raw images to fluorescence signal intensity, significantly improving the automation and accuracy of gene chip data analysis, while enhancing the adaptability and stability of the method under different experimental conditions.

[0005] Technical Solution: A deep learning-based method for analyzing fluorescence signals in gene chips. Its core lies in constructing a complete deep learning framework that can automatically learn effective features from gene chip images and accurately analyze fluorescence signal intensity. The method comprises four main stages: image acquisition and preprocessing, deep learning model construction, model training and optimization, and signal analysis and post-processing.

[0006] In the image acquisition and preprocessing stage, raw fluorescence images of the gene chip are acquired using specialized scanning equipment and then standardized. Preprocessing includes, but is not limited to, image denoising, contrast enhancement, and color balancing, aiming to eliminate the impact of equipment differences and environmental factors on image quality and provide high-quality input data for subsequent analysis. Specific preprocessing strategies are employed to ensure the clarity and identifiability of the microbead regions, taking into account the characteristics of gene chip images.

[0007] In the deep learning model construction phase, a complete architecture was designed, comprising a feature extraction module, a feature fusion module, and a signal analysis module. The feature extraction module employs a deep convolutional neural network structure, extracting feature representations at different abstract levels from the image through multi-layer convolutional operations. This module can automatically learn effective features related to fluorescence signals, avoiding the subjectivity and limitations of traditional methods that rely on manually designed features. The feature fusion module is responsible for integrating feature information from different network layers, enhancing the model's ability to utilize multi-scale features through features reweighting and feature concatenation. The signal analysis module performs quantitative analysis of fluorescence signals based on the extracted features, outputting the signal intensity value for each microbead region. The deep learning model employs a multi-task learning framework, simultaneously completing multiple related tasks such as microbead detection, classification, and signal intensity prediction. Through collaborative training and parameter sharing between tasks, the model's generalization ability and feature learning efficiency are improved. The model also introduces an attention mechanism, enabling the network to automatically focus on key regions in the image, improving the targeting and effectiveness of feature extraction.

[0008] In the model training and optimization phases, a supervised learning approach was adopted, using a dataset of accurately labeled gene chip images for model training. Advanced optimization algorithms and regularization techniques were employed during training to prevent overfitting and improve generalization performance. A specialized loss function was designed to address the characteristics of gene chip data, ensuring good prediction results on beads with varying brightness levels. Simultaneously, data augmentation techniques were used to expand the training samples, enhancing the model's robustness to image variations.

[0009] In the signal analysis and post-processing stage, the trained model is used to automatically analyze the new gene chip images. The system first locates all microbead regions in the image, then extracts features from each region and calculates the fluorescence intensity value. The post-processing process includes result verification, outlier correction, and data standardization to ensure the accuracy and reliability of the output results. The system also provides a visual interface for users to view the analysis results and make necessary adjustments.

[0010] This invention also provides a gene chip fluorescence signal analysis system based on deep learning. This system, implemented using the aforementioned method, includes an image acquisition module, a preprocessing module, a deep learning analysis module, and a result output module. The image acquisition module is responsible for acquiring the raw image data of the gene chip; the preprocessing module performs quality optimization and standardization on the images; the deep learning analysis module is the core of the system, containing a trained model capable of automatically performing feature extraction and signal analysis; and the result output module is responsible for generating analysis reports and visualization results.

[0011] Beneficial effects: (1) This invention automates the analysis of fluorescence signals in gene chips through deep learning technology, greatly reducing the need for manual intervention and improving analysis efficiency. Compared with traditional methods, this invention can quickly process large numbers of gene chip images, meeting the needs of high-throughput experiments.

[0012] (2) This method has stronger adaptability. Deep learning models can automatically learn effective feature representations from a large amount of data without the need for manual design of complex feature extraction rules. Therefore, it can adapt to gene chip images collected under different experimental conditions and with different equipment, and shows good robustness.

[0013] (3) The analysis method provided by this invention has higher accuracy. By using an end-to-end deep learning framework, the error accumulation problem that may be introduced by multiple processing steps in traditional methods is avoided. The model can capture subtle features related to fluorescence signals in the image and provide more accurate signal intensity quantification results.

[0014] (4) This method has good scalability. The deep learning-based framework can easily incorporate new network structures and technological advancements, and continuously improve system performance through continuous training and optimization. At the same time, the system can be seamlessly integrated with other bioinformatics analysis tools to form a complete gene chip data analysis workflow.

[0015] (5) The system provided by this invention is user-friendly. The entire analysis process is highly automated, and users only need to provide gene chip images to obtain accurate analysis results. The system also provides detailed analysis reports and a visual interface to help users understand and verify the analysis results.

[0016] (6) By introducing deep learning technology, this invention solves a series of problems existing in traditional gene chip fluorescence signal analysis methods, and provides a more intelligent, accurate and efficient analysis scheme. This method is not only applicable to gene expression analysis in scientific research, but also to multiple fields such as clinical diagnosis and environmental monitoring, and has broad application prospects and important practical value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example like Figure 1 As shown, this embodiment provides a deep learning-based method for analyzing fluorescence signals in gene chips. Its core lies in constructing a multi-module deep learning model with end-to-end prediction as its core. This model can automatically learn and extract multi-scale features related to fluorescence signals from the original scanned images of the gene chip, and simultaneously perform microbead localization, brightness / darkness classification, and accurate prediction of fluorescence signal intensity. The entire approach aims to overcome the inherent shortcomings of traditional methods, such as reliance on manual feature design, difficulty in adapting to different experimental conditions, and error accumulation due to segmented processing, thereby achieving automation, intelligence, and high precision in the gene chip signal analysis process.

[0020] First, in the image acquisition and preprocessing stage, the system acquires raw fluorescence image data using a high-resolution gene chip scanner. These images typically contain a large array of microbeads, each representing a specific gene probe. Due to factors such as experimental conditions, hybridization efficiency, scanning parameters, and environmental noise, the raw images often exhibit inconsistent quality, such as uneven brightness, low contrast, background noise, or color deviation. Therefore, a series of standardized preprocessing operations must be performed on the raw images to improve image quality and provide stable and reliable input for subsequent deep learning model analysis. The preprocessing workflow mainly includes image denoising, contrast enhancement, and color balancing. Image denoising employs adaptive filters, such as nonlocal mean denoising or wavelet threshold denoising, to effectively suppress random noise while preserving the edge and detail information of the microbeads. Contrast enhancement is achieved through a restricted adaptive histogram equalization method, aiming to enhance the distinction between the microbead region and the background, avoiding background noise amplification due to over-enhancement. Color balance processing primarily targets multi-channel fluorescence images. By referencing a standard color chart or the background area within the image, the intensity of each fluorescence channel is calibrated to ensure the comparability of image data acquired from different batches and devices. After preprocessing, the image is normalized to a uniform size and pixel value range to facilitate efficient model processing.

[0021] Next, we move on to the core aspect of deep learning model construction. This embodiment designs a deep learning network architecture centered on end-to-end prediction and comprising multiple functional modules. This architecture mainly consists of three core modules connected in series: a feature extraction module, a feature fusion module, and a signal parsing module. The feature extraction module is the foundation of the entire model; its function is to automatically learn and extract effective features relevant to fluorescence signal analysis from the preprocessed gene chip images. This invention preferably uses a deep convolutional neural network as the backbone of this module, such as mature architectures like ResNet, DenseNet, or EfficientNet, which have been pre-trained on large-scale image data, and fine-tunes them to adapt to the specific patterns of the gene chip images. This module gradually extracts feature representations from the original pixels at multiple abstract levels, from low-level edge and texture features to high-level semantic features, through stacked multi-layer convolutional layers, pooling layers, and activation function layers. The size and stride of the convolutional kernels are carefully designed to ensure effective capture of the morphology of the microbeads and the local details of their fluorescence signals. Compared with traditional fixed feature extraction operators, this data-driven approach can adaptively learn the most discriminative features for predicting fluorescence signal intensity, avoiding the limitations of manual design.

[0022] The feature fusion module is responsible for effectively integrating the feature information output from different network layers by the feature extraction module. Since the expression intensity of fluorescence signals in gene chip images varies, and their effective information has different importance at different scales, features from a single layer are often insufficient to comprehensively and accurately describe the state of microbeads. This module introduces a feature pyramid network or similar structure to fuse high-resolution, low-semantic features from shallow networks with low-resolution, high-semantic features from deep networks. Specifically, upsampling and lateral connections can be used to align deep and shallow feature maps in spatial dimensions, followed by fusion through channel concatenation or element-wise addition. To further enhance the model's ability to focus on key information, this module also integrates attention mechanisms, such as channel attention and spatial attention. Channel attention adaptively recalibrates the weights of each feature channel, allowing the model to focus more on information-rich channels; spatial attention learns a weight map to highlight spatial locations in the image that contribute more to the task, such as the central region of the microbeads. By combining multi-scale feature fusion with attention weighting, the model can comprehensively utilize information at different levels of abstraction and dynamically focus on key regions, thereby significantly improving the expressive power and robustness of features.

[0023] The signal analysis module is the part of the model that makes the final decision. It receives the fused and weighted feature maps and outputs specific analysis results. One of the innovations of this invention is the adoption of a multi-task learning framework. This module is designed to handle multiple closely related tasks simultaneously. These tasks mainly include: microbead localization (outputting bounding boxes through object detection or pixel-level masks through semantic segmentation to determine the precise region of each microbead), microbead brightness / darkness classification (classifying microbeads into categories such as "bright" or "dark," which is usually related to gene expression levels or hybridization states), and regression prediction of fluorescence signal intensity (outputting a continuous intensity value for each microbead). The advantage of multi-task learning is that by sharing the low-level feature extraction layer and partially sharing parameters between high-level task-specific layers, the model can learn more generalizable feature representations. For example, learning how to accurately locate microbeads helps the model extract features of that region more accurately for intensity prediction; while classifying microbeads by brightness / darkness provides a coarse-grained prior for the regression task, constraining the predicted range of intensity values. This synergy and constraint between tasks effectively improves the overall performance and stability of the model.

[0024] In the model training and optimization phases, this embodiment employs supervised learning. A precisely labeled gene chip image dataset needs to be constructed. The images in this dataset must be labeled using expert or gold-standard methods, with annotation information including the bounding box or pixel-level mask for each microbead, brightness / darkness category labels, and actual fluorescence signal intensity values. To improve the model's generalization ability and prevent overfitting, data augmentation techniques are extensively used during training. These techniques include, but are not limited to: random rotation, flipping, scaling, random adjustment of brightness and contrast, and adding Gaussian noise, to create more diverse training samples, enabling the model to adapt to various possible image variations. The goal of model training is to minimize a carefully designed composite loss function. This loss function is a weighted sum of multiple task loss terms, combining the loss from localization tasks (such as the loss for bounding box regression or segmentation), the loss from classification tasks (such as the cross-entropy loss for microbead brightness / darkness classification), and the loss from regression tasks (such as the smoothed L1 loss for fluorescence signal intensity prediction, which is less sensitive to outliers). Through this composite loss function, the model can achieve good performance on multiple tasks simultaneously. The optimization process employs an adaptive moment estimation optimizer combined with a learning rate decay strategy. Simultaneously, regularization methods such as Dropout and weight decay are introduced during training to further enhance the model's generalization ability.

[0025] In the signal analysis and post-processing stage, the trained model is deployed to automate the analysis of newly acquired gene chip images. After the system loads the model, the preprocessed image to be analyzed is input, and the model outputs preliminary results in an end-to-end manner, including the location, category, and predicted fluorescence intensity values ​​of all microbeads. Since the model predictions may contain individual outliers or errors, a post-processing workflow is needed to refine the results. Post-processing includes: using a non-maximum suppression algorithm to remove duplicate microbead detection boxes, ensuring that each microbead is counted only once; performing statistical tests on the predicted fluorescence intensity to identify and correct possible outliers; and finally, performing batch effect correction and standardization on the fluorescence intensity data of the entire chip based on experimental internal parameters or control samples, making the results comparable between different experiments. The system ultimately generates a detailed analysis report, which includes the final fluorescence intensity value of each microbead, overall chip quality assessment indicators (such as signal-to-noise ratio, microbead detection rate, etc.), and a visual result graph, such as using different colors to mark microbeads of different intensities or categories overlaying them on the original image, allowing users to intuitively verify the reliability of the analysis results.

[0026] like Figure 2 As shown, this embodiment also provides a gene chip fluorescence signal analysis system based on deep learning. The system integrates a high-performance graphics processing unit (GPU) computing server, an image acquisition device (such as a chip scanner), and a user interaction terminal in terms of hardware. At the software level, the system mainly includes the following functional modules: an image acquisition module, responsible for controlling the scanner hardware, acquiring raw fluorescence image data, and completing preliminary data format conversion and storage; a preprocessing module, receiving the raw image, executing the aforementioned series of image quality optimization and standardization algorithms, and outputting a standard format image that meets the model input requirements; a deep learning analysis module, the core of the system, which incorporates a pre-trained deep learning model based on end-to-end prediction. This module receives the preprocessed image, automatically completes the entire calculation process of feature extraction, feature fusion, and signal analysis using the model, and outputs preliminary microbead information and fluorescence intensity; and a results output module, responsible for post-processing and validating the preliminary model results, generating a final analysis report, and presenting the results to the user in the form of tables, charts, and visualizations. Users can easily upload data, start the analysis process, view and export results through the graphical user interface. The entire system is designed with a modular structure, which facilitates maintenance and upgrades. For example, when there are new deep learning models or algorithms, the deep learning analysis module can be easily replaced or updated.

[0027] This embodiment successfully achieves high-precision, automated analysis of gene chip fluorescence signals by constructing a deep learning model that integrates multiple modules and multi-task learning, with end-to-end prediction as its core. This method effectively overcomes many bottlenecks of traditional techniques, providing a powerful tool for gene expression analysis, clinical diagnosis, and related biotechnology fields.

[0028] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for analyzing a gene chip fluorescence signal based on deep learning, characterized by, The method comprises the following steps: S1, obtaining the original fluorescence image of the gene chip and performing standardization preprocessing, the preprocessing comprising image denoising, contrast enhancement and color balance; S2, constructing a deep learning model with end-to-end prediction as the core, the model comprising a feature extraction module, a feature fusion module and a signal analysis module in sequence; S3, using a gene chip image data set containing accurate labels to train and optimize the deep learning model in a supervised learning manner; S4, using the trained model to automatically locate the microbeads, analyze the signals and perform post-processing on the new gene chip image, and output the fluorescence signal intensity value of each microbead region. 2.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The feature extraction module adopts a deep convolutional neural network structure and automatically extracts feature representations of different abstraction levels related to the fluorescence signal from the image through multiple convolution operations. 3.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The feature fusion module is responsible for integrating feature information from different network layers in the feature extraction module, and realizes the fusion of multi-scale features through feature reweighting and feature splicing. 4.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The signal analysis module performs quantitative analysis of the fluorescence signal based on the features output by the feature fusion module; and the deep learning model adopts a multi-task learning framework and can simultaneously perform multiple tasks such as microbead positioning, bright-dark classification and signal intensity prediction. 5.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The microbead positioning task is realized by outputting a bounding box through target detection or outputting a pixel-level mask through semantic segmentation. 6.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The multi-task learning framework improves the generalization ability of the model through collaborative training and parameter sharing between tasks, and adopts a composite loss function combining classification loss and regression loss for optimization. 7.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The model introduces an attention mechanism, enabling the network to automatically focus on the key microbead regions in the image and dynamically adjust the feature weights. 8.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The model training and optimization stage adopts data enhancement techniques and regularization methods. 9.The deep learning-based gene chip fluorescence signal analysis method of claim 1, wherein, The signal analysis and post-processing stage comprises locating the microbead regions in the image and calculating the fluorescence intensity using the trained model, and then verifying the preliminary results, correcting outliers and performing data standardization processing.

10. A system using the deep learning-based gene chip fluorescence signal analysis method according to any one of claims 1 to 9, characterized in that, It comprises: An image acquisition module for obtaining original image data of a gene chip; A preprocessing module for quality optimization and standardization processing of the original image; A deep learning analysis module, which is the core processing unit, has a deep learning model trained according to the method of any one of claims 1-9 built in, and is used to automatically complete feature extraction, fusion and signal analysis; A result output module for generating an analysis report and visual results containing fluorescence signal intensity values.

Images