Single molecule image enhancement method based on adaptive deep learning

Abstract

The invention discloses a single-molecule fluorescence image denoising and quantifying method based on adaptive deep learning (ADeL-SM), and aims to solve the problems that the signal-to-noise ratio of an image is low in a single-molecule imaging process, quantitative information is damaged by a traditional denoising method and the like. According to the method, on the premise that external training data is not needed, a pseudo training pair is constructed through image sequence time averaging, a fluorescence intensity keeping mechanism and a self-feedback optimization strategy are introduced based on a neural network of an improved U-Net structure, a deep neural network model is trained, and image denoising enhancement is achieved. According to the method, the signal-to-noise ratio and positioning precision of the single-molecule image can be remarkably improved, quantitative information of fluorescence intensity is reserved, and the method is suitable for various single-molecule imaging platforms and biological research scenes and can be widely applied to biological quantitative imaging tasks such as membrane protein positioning, aggregation recognition and trajectory analysis.

Description

Technical Field

[0001] This invention relates to the intersection of single-molecule imaging, image processing, and artificial intelligence technologies, and in particular to a method for denoising and quantitatively enhancing single-molecule fluorescence images based on adaptive deep learning. Background Technology

[0002] Single-molecule fluorescence imaging has become an important tool for studying cell membrane protein distribution, receptor aggregation behavior, and molecular dynamics. However, in practical experiments, low imaging light intensity and high background noise result in poor signal-to-noise ratios in the original images, affecting the accuracy of molecular localization and subsequent analysis. Traditional image denoising methods, such as Gaussian filtering and median filtering, can remove noise to some extent, but they often damage the molecular signal itself, leading to loss of fluorescence intensity and localization errors.

[0003] Existing deep learning denoising methods mostly rely on large amounts of high-quality paired training data, which are difficult to obtain in single-molecule imaging experiments. Furthermore, their optimization goals are usually to preserve the image structure, neglecting the quantification of fluorescence intensity, thus limiting their application in quantitative biophysical research such as molecular trajectory tracking and diffusion coefficient calculation.

[0004] Against the backdrop of the aforementioned technologies, this invention proposes a deep learning-based single-molecule image denoising method that can adaptively generate training pairs, maintain fluorescence intensity, and improve the signal-to-noise ratio without relying on external training data. This method achieves high-precision single-molecule localization and quantitative analysis through an innovative self-feedback mechanism and intensity preservation strategy, significantly enhancing the application effects and biological research value of single-molecule fluorescence imaging. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning image denoising and enhancement method (ADeL-SM) that can adaptively construct training data pairs and preserve fluorescence intensity information without the need for external training data, thereby overcoming the limitations of existing methods in terms of image quantification and generalization.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] (1) Image acquisition: Single-molecule fluorescence microscopy equipment is used to acquire single-molecule fluorescence image sequences to be processed, ensuring that the acquired images contain rich spatiotemporal information and molecular dynamic characteristics.

[0008] (2) Training Pair Construction: To address the limitation of lacking external labeled data, a weighted time averaging method is used to effectively suppress random noise from multiple frames of images in the time series. Then, localization is performed, and the intensity of the localization is obtained from the feedback of the original image. The background is the minimum intensity in the original image, generating a pseudo-clean image as a reference image in the training pair. The original single-frame image is paired with the corresponding pseudo-clean image to form the input and output samples required for network training, avoiding dependence on real noise-free images.

[0009] (3) This invention constructs a convolutional neural network structure based on U-Net to enhance and quantitatively recover weak signals in single-molecule fluorescence images. This network optimizes the traditional U-Net symmetric encoder-decoder architecture: on the one hand, residual connections are introduced to alleviate the gradient vanishing problem in deep network training and improve model stability; on the other hand, skip connections are used to fuse shallow features in the encoder with high-level features in the decoder, effectively preserving edge and detail information of the image, thereby improving the spatial resolution and structural fidelity of the reconstructed image. Furthermore, the network input is normalized to unify the data scale, and the output employs an inverse normalization strategy to ensure high quantitative consistency of the recovered image at the physical light intensity level.

[0010] (4) In the current training process of this invention, the root mean square error (RMSE) loss function is used as the main loss term to measure the overall pixel intensity difference between the network output image and the reference image, thereby effectively reducing background noise and improving image restoration quality. To monitor the convergence and stability of the network during training, the curve of the loss function changing with the training epoch was plotted and analyzed.

[0011] (5) Application and output: The trained model can efficiently denoise any acquired single-molecule fluorescence image sequence, output image data with significantly improved signal-to-noise ratio and quantitatively maintained fluorescence intensity, support subsequent molecular localization, trajectory tracking and dynamic analysis, and significantly improve the quantitative accuracy of single-molecule imaging and biological research value.

[0012] The method of the present invention has the following advantages:

[0013] This invention offers significant advantages over existing technologies. First, it eliminates the need for external high-quality training data, exhibiting excellent experimental adaptability and applicability to various imaging conditions. Second, by introducing a self-feedback algorithm, it effectively maintains the intensity of molecular fluorescence signals, achieving high fidelity in quantitative information and significantly enhancing the model's robustness and generalization ability, enabling stable operation under diverse data environments. Finally, this invention significantly improves the signal-to-noise ratio and molecular localization accuracy of images, greatly enhancing the analytical reliability and biological research value of single-molecule fluorescence imaging. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall framework of the ADeL-SM system;

[0015] Figure 2 This is a schematic diagram of the construction of the ADeL-SM pseudo-training pair;

[0016] Figure 3 This is a schematic diagram of the network structure;

[0017] Figure 4 The graph shows (a) the root mean square error loss and (b) the loss function curve;

[0018] Figure 5 The distribution maps are (a) the original image and (b) the enhanced image, and (c) the signal-to-noise ratio and (d) the positioning accuracy.

[0019] Figure 6 The images show the original and enhanced (a) single-molecule trajectory diagrams and (b) intensity correlation diagrams.

[0020] Figure 7 The images show the original and enhanced monomolecules: (a) step bleaching diagram and (b) motion trajectory diagram. Detailed Implementation

[0021] This invention addresses the challenges of low signal-to-noise ratio and quantification in single-molecule fluorescence imaging by proposing an image denoising and quantification method based on adaptive deep learning. Through innovative training pair construction and self-feedback mechanisms, fluorescence intensity fidelity is ensured, effectively improving image quality and molecule localization accuracy, meeting the requirements of high-precision single-molecule analysis. The invention is further illustrated below with reference to the accompanying drawings and embodiments: Embodiment 1

[0022] (1) Image acquisition

[0023] like Figure 1As shown, the method of this invention uses total internal reflection fluorescence microscopy (TIRF) to image the sample to obtain a sequence of single-molecule fluorescence images on the cell membrane surface. This sequence contains 300 consecutive images, reflecting the spatial distribution of single molecules on the cell membrane and their dynamic changes over time. The imaging conditions are an excitation wavelength of 488 nm and an exposure time of 100 ms to ensure maximum acquisition of single-molecule fluorescence signals while minimizing the effects of photobleaching and background noise. The acquired images are two-dimensional grayscale images, with resolution and sampling rate meeting the accuracy requirements for single-molecule detection and trajectory tracking.

[0024] (2) Training on the construction

[0025] like Figure 2 As shown, to address the practical problems of low signal-to-noise ratio and difficulty in obtaining high-quality labeled data for single-molecule fluorescence images, this method uses two consecutive frames from a time series image. The pixel intensities are time-weightedly averaged to suppress random noise while preserving the stable portion of the single-molecule signal. ThunderSTORM is used to locate the two averaged frames. The intensity of the 5×5 pixel location points is assigned the same as the original image intensity, while the remaining intensities are assigned the lowest intensity from the original image. The resulting pseudo-clean image forms a training data pair with the corresponding original single-frame image for network supervised training. This method avoids the need for real, noise-free images, adapts to different experimental conditions, and ensures the diversity and authenticity of the training data.

[0026] (3) Construction and training of deep learning networks

[0027] Figure 3 This paper illustrates the deep convolutional neural network structure used in this invention. Based on the U-Net structure, this network consists of an encoder and a decoder. The encoder extracts image features layer by layer through multi-layer convolution and downsampling operations; the decoder restores the spatial resolution of the image through progressive upsampling. To alleviate the gradient vanishing problem that may occur during the training of deep networks, residual connections are introduced to enhance the fluidity of feature information between layers. Simultaneously, skip connections are preserved and strengthened, allowing high-resolution detail features from shallow layers to be fused with semantic information from deeper layers, effectively improving the ability to recover edges and details. The loss function obtained during training in this embodiment is... Figure 4 Root Mean Square Error (RMSE) and Figure 4 The loss function results, RMSE and loss, are basically stable after training for 2 epochs, indicating that the model has good training effect and can be used for image prediction.

[0028] (4) Model application and image enhancement

[0029] After completing the training, such as Figure 5 As shown, the obtained model is applied to new single-molecule fluorescence image sequences, for Figure 5 The original image is subjected to Adele-SM denoising and enhancement processing to obtain Figure 5 b. This model can significantly improve the image signal-to-noise ratio. Figure 5 c) Improves positioning accuracy by 6.4 times, highlighting single-molecule signals, reducing background noise interference, and enhancing positioning accuracy. Figure 5 d) Increased fluorescence intensity by 1.6 times while maintaining quantitative accuracy. The enhanced images have higher resolution and localization precision, laying the foundation for subsequent molecular recognition and dynamic analysis. Figure 6 a original image and Figure 6 Fluorescence intensity in the single-molecule scintillation curve of the bADeL-SM enhanced image was measured. Figure 6 In terms of intensity correlation, the Pearson correlation coefficient was calculated to be 0.9856. The single-molecule fluorescence intensity was preserved before and after image enhancement, while the background was significantly reduced. Furthermore, the model exhibits strong adaptability, applicable to image data acquired under different samples and experimental conditions, greatly improving the practicality and reliability of single-molecule imaging.

[0030] (6) Subsequent molecular analysis

[0031] In addition, combined Figure 7 This invention also supports single-molecule dynamic behavior analysis based on ADeL-SM enhanced images. By extracting trajectories and performing kinematic analysis on the time series of localized molecules, higher accuracy calculations can be achieved. Figure 7 a clustered distribution, a more complete reconstruction Figure 7 Key biophysical parameters such as β-molecule diffusion provide important data support for revealing the function of cell membrane proteins and their mechanisms in immune escape and drug action.

[0032] In summary, this invention proposes a single-molecule fluorescence image enhancement method that integrates pseudo-training pair construction, self-feedback mechanism, and deep network optimization. This method requires no external training data, maintains fluorescence intensity fidelity, exhibits high adaptability and enhancement effect, significantly improves the quantitative analysis quality of single-molecule images, and has significant scientific research value and broad application prospects. Although the invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the invention fall within the scope of protection claimed by this invention.

Claims

1. A single-molecule image enhancement method based on adaptive deep learning, characterized in that... The specific steps are as follows: Claim 1: A method for denoising and quantitative analysis of single-molecule fluorescence images based on adaptive deep learning, characterized by comprising the following steps: acquiring image sequences, generating pseudo-training pairs, training a deep neural network, introducing an intensity preservation mechanism and a self-feedback structure, and outputting high signal-to-noise ratio images for subsequent quantitative analysis.

2. Claim 2: The method according to claim 1, characterized in that, The training pair is constructed by combining the time-averaged image with the original image, with a frame count of no less than 100 frames.

3. Claim 3: The method according to claim 1, characterized in that, The neural network is an improvement on the U-Net structure and has residual connections.

4. Claim 4: The method according to claim 1, characterized in that, Root mean square error (RMSE) and loss maintain accuracy.

5. Claim 5: The method according to claim 1, characterized in that, This method is applicable to total internal reflection fluorescence microscopes, confocal microscopes, and other microscope platforms.

6. Claim 6: The method according to claim 1, characterized in that, The output images are used for the quantification of membrane protein distribution, diffusion, and aggregation states.

Images