Medical image focus area positioning method based on deep learning

Abstract

The invention relates to the technical field of medical image recognition, and provides a medical image focus area positioning method based on deep learning, which comprises the following steps: acquiring a to-be-detected medical image; performing format unification and size unification on the to-be-measured medical image to obtain a standardized image; denoising the standardized image to obtain a denoised image; the de-noised image is converted into an RBG image; inputting the RBG image into a trained focus area positioning model to obtain a positioning result; the lesion area positioning model is constructed based on a U-Net network, and the feature of the U-Net enables the lesion area positioning model to better capture fine features, so that the problem of boundary blur is reduced. And random noise and artifacts in the image can be eliminated by denoising the standardized image, so that extraction and recognition of disease features are facilitated, and the segmentation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image recognition, and in particular to a method for locating a medical image lesion region based on deep learning. Background Art

[0002] In recent years, with the development of medical imaging technology, various medical imaging devices such as X-ray, computed tomography (CT), and magnetic resonance imaging (MRI) have been widely used in clinical diagnosis. These imaging technologies can provide high-resolution images of the internal structure of the human body, providing important support for the early diagnosis and treatment of diseases. However, manual analysis of large amounts of medical imaging data is not only time-consuming and labor-intensive, but also easily affected by expert experience and fatigue, resulting in inconsistent diagnostic results. Therefore, how to use advanced computer technology to efficiently and accurately automatically identify and locate lesion areas in medical images has become a hot topic in current research.

[0003] Deep learning, especially convolutional neural networks (CNNs), has gradually become an important tool in the field of medical image analysis due to its excellent performance in image processing and pattern recognition. Deep learning models have the ability to automatically learn image features and can accurately locate complex and hidden lesions. Compared with traditional image processing methods, deep learning can improve the robustness and recognition accuracy of the model by processing a large amount of annotated data. In addition, with the improvement of hardware computing power and the accumulation of massive medical data, the application of deep learning in medical image analysis has great potential.

[0004] Application No. 202010651956.3 discloses a method for classification and segmentation of lung pathological images. The specific method is: input a full-slice pathological image; use a sliding window to segment the pathological image to obtain image blocks; use a lesion type classification model to analyze the foreground image blocks in turn, and identify the lesion type of the tissue area in the foreground image block; output the lesion type classification result; splice the foreground image block lesion area segmentation result according to the relative position of the corresponding foreground image block in the pathological image, and fill the background with the background image block area, and obtain the pathological image lesion area segmentation result. This patent realizes the accurate identification of the lesion type of the lung pathological image while segmenting the specific boundaries of the lesion area. However, the segmentation result of the lesion area of ​​the patent has the problem of unclear boundaries, especially when the transition between the lesion and normal tissue is not obvious, which affects the doctor's judgment. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a medical image lesion area localization method based on deep learning, which can reduce the problem of blurred boundaries and enhance the segmentation accuracy of medical images.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for locating lesion areas in medical images based on deep learning, comprising:

[0008] Acquire the medical image to be tested;

[0009] Unifying the format and size of the medical image to be tested to obtain a standardized image;

[0010] Denoising the standardized image to obtain a denoised image;

[0011] Converting the denoised image into an RBG image;

[0012] The RBG image is input into a trained lesion area localization model to obtain a localization result; the lesion area localization model is constructed based on a U-Net network.

[0013] Preferably, denoising the standardized image to obtain a denoised image comprises:

[0014] Using a filter window to detect noise points on the standardized image, to obtain noise points to be processed;

[0015] When the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising the standardized image in the corresponding filtering window;

[0016] Slide the filter window and return to the step of "using the filter window to detect noise points on the standardized image to obtain noise points to be processed" until the entire standardized image is traversed to obtain a filtered image;

[0017] Take a neighborhood window centered at any point on the filtered image and calculate the grayscale average of all pixels in the neighborhood window;

[0018] The grayscale average value of the corresponding pixel points is taken as the output of the central pixel point to obtain the denoised image.

[0019] Preferably, using a filter window to detect noise points on the standardized image to obtain noise points to be processed includes:

[0020] A noise point detection model is constructed based on the mean and median of each image point in the filtering window;

[0021] Detecting each image point in the filtering window using the noise point detection model to obtain a similar noise value for each image point;

[0022] The corresponding image points that are larger than the similar noise value are taken as noise points to be processed.

[0023] Preferably, the noise point detection model is:

[0024]

[0025] Among them, f(x) represents the similar noise value of pixel x, u(x) represents the gray value of pixel x, and u mean (x) represents the grayscale mean of all pixels in the filter window centered on pixel x, ▽u(x) represents the gradient mean of pixel x, ▽u mean (x) is the grayscale median of all pixels in the filter window centered on pixel x, ▽ x (x) represents the gradient value of pixel x in the horizontal direction, ▽ y (x) represents the gradient value of pixel x in the vertical direction.

[0026] Preferably, when the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising is performed on the standardized image in the corresponding filtering window, including:

[0027] The pseudo pixel variance is calculated according to the grayscale median of all pixels in the filter window; wherein the pseudo pixel variance calculation formula is: in, represents the pseudo pixel variance of the pixel point (a, b) in the area of ​​the filter window with a size of (2n+1)×(2n+1), mean(a,b) represents the grayscale median of the pixel point (a,b) in the filter window, and x(k,l) represents the grayscale value of the pixel point at the position (k,l);

[0028] A window denoising model is constructed using the pseudo pixel variance.

[0029] Preferably, constructing a window denoising model using the pseudo pixel variance includes:

[0030] Using the formula:

[0031]

[0032] Construct a window denoising model; where f(a, b) represents the grayscale value of the pixel (a, b) after denoising, D is an adjustable coefficient, and x(a, b) represents the grayscale value of the pixel (a, b) in the filter window.

[0033] Preferably, converting the denoised image into an RBG image comprises:

[0034] Obtaining the grayscale value of each pixel on the denoised image;

[0035] Determining a conversion threshold according to the grayscale value of each pixel on the denoised image;

[0036] A RBG image conversion model is constructed according to the conversion threshold, and the denoised image is converted into an RBG image using the RBG image conversion model; wherein the RBG image conversion model is:

[0037]

[0038] Among them, Gmin represents the minimum grayscale value of the denoised image, Gmax represents the maximum grayscale value of the denoised image, Thre1 represents the first conversion threshold, and Thre1=Gmin+(Gmax-Gmin) / 3, f(x, y) represents the grayscale value of the image at the corresponding point, Thre2 represents the second conversion threshold, and Thre2=Gmax-(Gmax-Gmin)*3, k1, k2, k3 represent adjustable parameters, and R, G and B respectively represent the values ​​of each image channel in the RBG image.

[0039] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] The present invention provides a method for locating a medical image lesion region based on deep learning, comprising: obtaining a medical image to be tested; unifying the format and size of the medical image to be tested to obtain a standardized image; denoising the standardized image to obtain a denoised image; converting the denoised image into an RBG image; inputting the RBG image into a trained lesion region positioning model to obtain a positioning result; the lesion region positioning model is constructed based on a U-Net network, and the characteristics of the U-Net of the present invention enable it to better capture subtle features, thereby reducing the problem of blurred boundaries. And the present invention denoises the standardized image to eliminate random noise and artifacts in the image, which will help extract and identify disease features, thereby improving the accuracy of segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1 A flow chart of a method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The purpose of the present invention is to provide a method for locating the lesion area in medical images based on deep learning. The characteristics of U-Net enable it to better capture subtle features, thereby reducing the problem of blurred boundaries. In addition, the present invention denoises the standardized image to eliminate random noise and artifacts in the image, which will help extract and identify disease features, thereby improving the accuracy of segmentation.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a method for locating a medical image lesion region based on deep learning, comprising:

[0047] Step 100: Acquire the medical image to be tested;

[0048] Step 200: unifying the format and size of the medical image to be tested to obtain a standardized image;

[0049] Step 300: De-noising the standardized image to obtain a de-noised image;

[0050] Step 400: converting the denoised image into an RBG image;

[0051] Step 500: input the RBG image into a trained lesion area localization model to obtain a localization result; the lesion area localization model is constructed based on a U-Net network.

[0052] Specifically, this embodiment collects multimodal medical imaging data, including MRI, CT, X-Ray, etc.

[0053] Optionally, the specific steps of format unification in this embodiment are as follows:

[0054] Convert medical images from raw formats (such as DICOM, TIFF, JPEG, etc.) to a unified image format, such as converting all images to JPEG or PNG format for subsequent processing.

[0055] Furthermore, the specific steps of size unification in this embodiment are as follows:

[0056] An appropriate target size (such as 256x256 pixels, 512x512 pixels, etc.) is selected, and all images are adjusted to the size through a scaling operation using an interpolation method (such as bilinear interpolation or cubic interpolation) to ensure image quality.

[0057] For images whose size is larger than the target size, center cropping is performed to retain the center part of the image.

[0058] For images that are smaller than the target size, you can choose edge padding or zero padding to ensure that the images reach a uniform size.

[0059] Furthermore, the standardization processing steps of this embodiment are as follows:

[0060] The pixel values ​​of the image are normalized and the pixel value range is adjusted to [0, 1] or [-1, 1] to meet the input requirements of the deep learning model.

[0061] Preferably, denoising the standardized image to obtain a denoised image comprises:

[0062] Using a filter window to detect noise points on the standardized image, to obtain noise points to be processed;

[0063] When the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising the standardized image in the corresponding filtering window;

[0064] Slide the filter window and return to the step of "using the filter window to detect noise points on the standardized image to obtain noise points to be processed" until the entire standardized image is traversed to obtain a filtered image;

[0065] Take a neighborhood window centered at any point on the filtered image and calculate the grayscale average of all pixels in the neighborhood window;

[0066] The grayscale average value of the corresponding pixel points is taken as the output of the central pixel point to obtain the denoised image.

[0067] Preferably, using a filter window to detect noise points on the standardized image to obtain noise points to be processed includes:

[0068] A noise point detection model is constructed based on the mean and median of each image point in the filtering window;

[0069] Detecting each image point in the filtering window using the noise point detection model to obtain a similar noise value for each image point;

[0070] The corresponding image points that are larger than the similar noise value are taken as noise points to be processed.

[0071] Preferably, the noise point detection model is:

[0072]

[0073] Among them, f(x) represents the similar noise value of pixel x, u(x) represents the gray value of pixel x, and u mean (x) represents the grayscale mean of all pixels in the filter window centered on pixel x, ▽u(x) represents the gradient mean of pixel x, ▽u mean (x) is the grayscale median of all pixels in the filter window centered on pixel x, ▽ x (x) represents the gradient value of pixel x in the horizontal direction, ▽ y (x) represents the gradient value of pixel x in the vertical direction.

[0074] Specifically, this embodiment constructs a noise point detection model based on the mean, median and gradient mean of each image point in the filtering window, which can detect the difference between noise points and original pixel points from multiple aspects, thereby making the detection of noise points more accurate.

[0075] Furthermore, the original filtering algorithm of this embodiment, such as the mean filtering algorithm, performs mean processing on the pixels in each neighborhood of the image (regardless of whether there are noise points), so the processed image will become blurred, while the present invention can find the noise points on the image by using the noise point detection model, and then filter the corresponding noise points to smooth out the noise points in the image while maintaining the original pixel information of the image. In practical applications, the present invention can set the corresponding detection interval value according to the actual scenario.

[0076] Preferably, when the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising is performed on the standardized image in the corresponding filtering window, including:

[0077] The pseudo pixel variance is calculated according to the grayscale median of all pixels in the filter window; wherein the pseudo pixel variance calculation formula is: in, represents the pseudo pixel variance of the pixel point (a, b) in the area of ​​the filter window with a size of (2n+1)×(2n+1), mean(a,b) represents the grayscale median of the pixel point (a,b) in the filter window, and x(k,l) represents the grayscale value of the pixel point at the position (k,l);

[0078] A window denoising model is constructed using the pseudo pixel variance.

[0079] Preferably, constructing a window denoising model using the pseudo pixel variance includes:

[0080] Using the formula:

[0081]

[0082] Construct a window denoising model; where f(a, b) represents the grayscale value of the pixel (a, b) after denoising, D is an adjustable coefficient, and x(a, b) represents the grayscale value of the pixel (a, b) in the filter window.

[0083] Furthermore, the present invention is based on a window denoising model and denoises the image within the corresponding filter window, which can alleviate the problem that existing denoising methods (such as median filtering denoising, mean filtering denoising and wavelet denoising, etc.) cause certain characteristic gradients in the nuclear magnetic resonance image to disappear, and can retain the original pixel information of the image to the greatest extent, thereby improving the image interpretation effect.

[0084] For example, in the data preprocessing stage, this embodiment performs denoising on the standardized image to eliminate random noise and artifacts in the image, and this embodiment can extract and identify disease features, thereby improving the accuracy of segmentation. And by normalizing the input image format and size, it can ensure that the model receives data with high consistency, thereby improving the learning efficiency and accuracy of the network during segmentation.

[0085] Preferably, converting the denoised image into an RBG image comprises:

[0086] Obtaining the grayscale value of each pixel on the denoised image;

[0087] Determining a conversion threshold according to the grayscale value of each pixel on the denoised image;

[0088] A RBG image conversion model is constructed according to the conversion threshold, and the denoised image is converted into an RBG image using the RBG image conversion model; wherein the RBG image conversion model is:

[0089]

[0090] Among them, Gmin represents the minimum grayscale value of the denoised image, Gmax represents the maximum grayscale value of the denoised image, Thre1 represents the first conversion threshold, and Thre1=Gmin+(Gmax-Gmin)*3, f(x,y) represents the grayscale value of the image at the corresponding point, Thre2 represents the second conversion threshold, and Thre2=Gmax-(Gmax-Gmin)*3, k1, k2, k3 represent adjustable parameters, and R, G and B respectively represent the values ​​of each image channel in the RBG image.

[0091] Specifically, the present invention performs pseudo-color processing on the denoised image based on a smooth, nonlinear sine function, and can adjust the values ​​of k1, k2, and k3 according to actual needs, so that the color saturation of the processed image is richer, the texture and edge areas are more obvious, and the contrast and layering of the image are greatly improved, so that the subsequent segmentation effect of the image is greatly improved. In other words, the basic principle of this embodiment is to directly divide the image greater than a certain gray value into the target area, and the image less than a certain gray value into the background area. Since this embodiment converts the grayscale image into a color image, it can directly use the values ​​of each channel in the RBG image for segmentation. That is to say, this embodiment quantizes a single gray value into three RBG channel values, and then this embodiment constructs color weights, position weights, and vector weights based on the three RBG channel values. Based on these parameters, the image segmentation can be made more detailed.

[0092] Furthermore, the lesion area localization model based on the U-Net network in this embodiment is mainly used for medical image segmentation. U-Net is a fully convolutional neural network composed of a symmetrical encoder (downsampling path) and a decoder (upsampling path). The feature maps of the symmetrical layers are connected through skip connections to accurately locate and segment the target area in the image. Its construction method and structural design are as follows:

[0093] The U-Net network of this embodiment includes:

[0094] The encoder (downsampling path) is set with an input layer that takes a normalized RGB image (assuming a size of 256x256x3). The convolutional block (repeated several times, such as 4 times) includes a convolutional layer, two 3x3 convolutional layers, using ReLU as the activation function, and a max pooling layer, 2x2 max pooling with a stride of 2, to achieve feature downsampling. With each repetition of the convolutional block, the spatial size of the feature map is halved, but the number of channels is doubled. This extracts progressively more abstract and high-level features.

[0095] The bottleneck layer is used for deep feature extraction, including two consecutive layers of 3x3 convolution followed by ReLU activation function. The feature map reaches the minimum size and the number of channels reaches the maximum (such as 1024 channels).

[0096] The decoder (upsampling path) is equipped with an upsampling block (repeated several times, such as 4 times). The upsampling block includes: upsampling layer: use 2x2 deconvolution layer (upsampling layer) to enlarge the feature map size and restore the feature map size of each layer; jump connection, connect the upsampled feature map with the feature map of the corresponding encoder stage, and merge the features that are closer to the details of the original image; convolution layer, two layers of 3x3 convolution, increase the diversity of channels and reduce the number of channels.

[0097] The output layer includes convolutional layers, and the last layer uses 1x1 convolution to project the higher-dimensional feature map to the number of categories of the original input image. The activation function uses Softmax or Sigmoid (for binary classification) to generate a probability map to clarify the probability that the pixel is in the lesion area.

[0098] Furthermore, the encoding and decoding paths of this embodiment are symmetrical, and the output feature map of each encoding block is directly connected to the symmetrical position of the decoder through a long connection, providing fine spatial information. And this embodiment uses the change of the channel to control the expressiveness of the learning feature, gradually hiding and revealing the lesion information in the channel. In addition, with the help of the splicing operation, it is ensured that the lesion is accurately located at the boundary using the context information.

[0099] In summary, through this carefully designed U-Net structure, the model of this embodiment not only effectively processes context information, but also accurately retains spatial information, making the lesion area positioning shorter and faster.

[0100] Specifically, the structural design of the U-Net network of this embodiment can effectively extract and retain image features through the symmetry of the encoder and decoder, as well as skip connections. During the segmentation process, especially in locations where the transition between lesions and normal tissues is not obvious, the characteristics of U-Net enable it to better capture subtle features, thereby reducing the problem of blurred boundaries. In addition, this embodiment can also use deformable convolution, which can further enhance the model's ability to handle complex boundary morphology. Deformable convolution allows the model to flexibly adjust the shape of the convolution kernel, thereby making local adjustments based on the actual morphology of the lesion and improving the accuracy of segmentation.

[0101] Furthermore, the model training method of this embodiment is as follows:

[0102] 1. Data preparation and enhancement,

[0103] Collect medical image data sets related to the task, and perform preliminary cleaning and annotation. For example, this embodiment applies a variety of data enhancement techniques (such as rotation, scaling, flipping, cropping, brightness / contrast adjustment, etc.) to increase the diversity of training samples, which can improve the robustness and generalization ability of the model, especially the recognition ability of lesions with blurred boundaries and deformation.

[0104] 2. Transfer learning, using the U-Net model pre-trained on a large-scale medical image dataset (such as ImageNet) as the initial weights. By fine-tuning the existing network parameters, it quickly converges and improves performance on specific tasks. This method is particularly effective when data is insufficient because it can better utilize existing knowledge.

[0105] 3. Construct a hybrid loss function. In order to improve the accuracy of the segmentation boundary and deal with the class imbalance problem, this embodiment designs a hybrid loss function. The hybrid loss function combines Dice loss and cross entropy loss, where Dice loss is used to emphasize the overlapping area, prompting the model to pay more attention to the segmentation of the lesion area. Cross entropy loss is used to ensure the classification validity of the entire image so that the model does not ignore the background area. By adjusting the weights of the two losses, a more balanced training goal is achieved.

[0106] 4. Adaptive learning rate adjustment: Use an adaptive learning rate algorithm (such as Adam or RMSprop) to automatically adjust the learning rate according to the performance of the validation set during training. By monitoring the loss value or accuracy, the learning rate is gradually reduced when the validation set performance improves, helping the model to update weights more stably when it is close to convergence.

[0107] 5. Batch normalization and group normalization. During the model training process, batch normalization layers and group normalization layers are added after each convolution layer to ensure that the distribution of each layer input is more stable and accelerate the training process. Through the above method, this embodiment can effectively alleviate the gradient vanishing problem and improve the stability of the model, especially when processing high-dimensional medical images.

[0108] 5. Regular validation and early stopping strategy: evaluate the performance of the model on the validation set after each training cycle (epoch), and record indicators (such as Dice coefficient and IoU) to monitor the training process of the model. If the performance on the validation set does not improve within multiple cycles, the early stopping mechanism is triggered to prevent the model from overfitting and retain the best weights.

[0109] 6. Result visualization and analysis: After the training process is completed, use visualization techniques (such as Grad-CAM or feature map visualization) to analyze the model's decision-making process to ensure that the model has captured the key features. The above steps can help verify the model's effectiveness in the boundary fuzzy area and provide guidance for further optimization.

[0110] Specifically, this embodiment comprehensively improves the training efficiency and final effect of the U-Net network in the task of localizing the lesion area in medical images through steps such as data enhancement, transfer learning, hybrid loss, adaptive learning rate, standardization technology, verification monitoring and result visualization. The comprehensive application of the strategies in this embodiment will effectively improve the performance of the model in the identification of blurred boundaries and complex lesions, providing more reliable support for clinical applications.

[0111] Optionally, the formula of the hybrid loss function of this embodiment is:

[0112] L mixed =α·L Dice +(1-α)·L CrossEntropy

[0113] Among them, L mixed is a mixed loss function, which serves as the target loss value in the overall training process. Dice It is the Dice loss, which aims to maximize the overlap between the predicted segmentation and the true segmentation. Its formula can be expressed as:

[0114]

[0115] Among them, P i is the probability value of the i-th pixel predicted by the model, G i is the segmentation label (0 or 1) of the i-th pixel actually marked. ∈ is a small positive number used to avoid the denominator being zero during calculation, usually set to 1e-7 or 1e-10.

[0116] L CrossEntropy It is a cross-extraction loss, which is often used in classification tasks. Its formula can be expressed as:

[0117]

[0118] Among them, N is the total number of pixels, G i is the label (0 or 1) of the i-th pixel actually marked. i is the probability value of the i-th pixel predicted by the model.

[0119] α is a weight coefficient that controls the relative contribution of Dice loss and cross entropy loss in the mixed loss. This value is usually adjusted between 0 and 1, and the specific value can be selected according to the effect of the validation set. Exemplarily, this embodiment uses a dynamic learning rate adjustment strategy to determine the weight coefficient, that is, in each training cycle (epoch), the Dice coefficient or cross entropy loss on the validation set is monitored, and the value of α is adjusted according to the change of the indicator.

[0120] Furthermore, the implementation adjustment steps of this embodiment are as follows:

[0121] Initialize α to 0.5 (half / half contribution).

[0122] At the end of each epoch, the loss and performance evaluation indicators (such as the Dice coefficient) of the validation set are calculated.

[0123] Adjust α according to the improvement of the indicators: If the Dice coefficient of the validation set increases, increase α appropriately (such as adjusting it to 0.6 or higher). If the cross entropy loss remains high, reduce α appropriately (such as adjusting it to 0.4 or lower). In addition, this embodiment can also set some restrictions to prevent α from exceeding the range of [0,1]. This embodiment combines Dice loss and cross entropy loss, complementing each other's strengths and weaknesses, and aims to simultaneously improve the overlap detection capability and classification accuracy of the model. This method can effectively deal with the class imbalance problem in medical image segmentation and enhance the model's understanding of the lesion area and background part.

[0124] In summary, the design of the medical image lesion area localization method based on deep learning in this embodiment can effectively deal with the boundary blur and filling problems mentioned in the prior art through the U-Net architecture and other deep learning technologies. The corresponding denoising, standardization and intelligent filling strategies will further improve the performance of the model in this regard, thereby providing doctors with more accurate and well-visualized lesion localization results.

[0125] The beneficial effects of the present invention are as follows:

[0126] (1) The design of the hybrid loss function of the present invention enables the model to pay more attention to the lesion area during the learning process, thereby improving the accuracy of segmentation and the ability to capture tiny features in the image.

[0127] (2) The mechanism of dynamically adjusting α of the present invention enables the model to reasonably balance overlap and classification losses at different training stages, adapt to the complexity of the segmentation task, and improve the stability and effectiveness of the model.

[0128] (3) The present invention uses an efficient denoising algorithm and accurate noise point detection. The model can remove noise while retaining the original information of the image to the maximum extent, thereby improving the subsequent segmentation effect and diagnostic accuracy.

[0129] (4) The present invention can process a variety of image data formats and diverse medical image data, enhance the versatility of the model, provide rich support for clinical applications, and provide solutions for different types of medical image segmentation tasks.

[0130] (5) The present invention utilizes pre-trained models and transfer learning to achieve fast and effective training, greatly reducing training time and cost, and is suitable for rapid application and iteration in clinical environments.

[0131] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0132] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for locating lesion areas in medical images based on deep learning, characterized in that: include: Acquire the medical image to be tested; Unifying the format and size of the medical image to be tested to obtain a standardized image; Denoising the standardized image to obtain a denoised image; Converting the denoised image into an RBG image; The RBG image is input into a trained lesion area localization model to obtain a localization result; the lesion area localization model is constructed based on a U-Net network.

2. The method for locating medical image lesion regions based on deep learning according to claim 1, characterized in that: Denoising the standardized image to obtain a denoised image, comprising: Using a filter window to detect noise points on the standardized image, to obtain noise points to be processed; When the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising the standardized image in the corresponding filtering window; Slide the filtering window and return to step "using the filtering window to detect noise points on the standardized image to obtain noise points to be processed" until the entire standardized image is traversed to obtain a filtered image; Take a neighborhood window centered at any point on the filtered image and calculate the grayscale average of all pixels in the neighborhood window; The grayscale average value of the corresponding pixel points is taken as the output of the central pixel point to obtain the denoised image.

3. The method for locating medical image lesion areas based on deep learning according to claim 2, characterized in that: Using a filter window to detect noise points on the standardized image to obtain noise points to be processed includes: A noise point detection model is constructed based on the mean and median of each image point in the filtering window; Detecting each image point in the filtering window using the noise point detection model to obtain a similar noise value for each image point; The corresponding image points that are larger than the similar noise value are taken as noise points to be processed.

4. The method for locating medical image lesion regions based on deep learning according to claim 3, characterized in that: The noise point detection model is: Among them, f(x) represents the similar noise value of pixel x, u(x) represents the gray value of pixel x, and u mean (x) represents the grayscale mean of all pixels in the filter window centered on pixel x, ▽u(x) represents the gradient mean of pixel x, ▽u mean (x) is the grayscale median of all pixels in the filter window centered on pixel x, ▽ x (x) represents the gradient value of pixel x in the horizontal direction, ▽ y (x) represents the gradient value of pixel x in the vertical direction.

5. The method for locating medical image lesion regions based on deep learning according to claim 4, characterized in that: When the number of noise points to be processed in the filtering window is greater than a preset threshold, denoising is performed on the standardized image in the corresponding filtering window, including: The pseudo pixel variance is calculated according to the grayscale median of all pixels in the filter window; wherein the pseudo pixel variance calculation formula is: in, represents the pseudo pixel variance of the pixel point (a, b) in the area of ​​the filter window with a size of (2n+1)×(2n+1), mean(a,b) represents the grayscale median of the pixel point (a,b) in the filter window, and x(k,l) represents the grayscale value of the pixel point at the position (k,l); A window denoising model is constructed using the pseudo pixel variance.

6. The method for locating medical image lesion regions based on deep learning according to claim 5, characterized in that: The pseudo pixel variance is used to construct a window denoising model, including: Using the formula: Construct a window denoising model; where f(a, b) represents the grayscale value of the pixel (a, b) after denoising, D is an adjustable coefficient, and x(a, b) represents the grayscale value of the pixel (a, b) in the filter window.

7. The method for locating medical image lesion regions based on deep learning according to claim 6, characterized in that: Converting the denoised image into an RBG image includes: Obtaining the grayscale value of each pixel on the denoised image; Determining a conversion threshold according to the grayscale value of each pixel on the denoised image; A RBG image conversion model is constructed according to the conversion threshold, and the denoised image is converted into an RBG image using the RBG image conversion model; wherein the RBG image conversion model is: Among them, Gmin represents the minimum grayscale value of the denoised image, Gmax represents the maximum grayscale value of the denoised image, Thre1 represents the first conversion threshold, and Thre1=Gmin+(Gmax-Gmin) / 3, f(x, y) represents the grayscale value of the image at the corresponding point, Thre2 represents the second conversion threshold, and Thre2=Gmax-(Gmax-Gmin) / 3, k1, k2, k3 represent adjustable parameters, and R, G and B respectively represent the values ​​of each image channel in the RBG image.

Images