Method for automatic localization in medical image based on deep learning

Abstract

The present invention relates to the technical field of automatic localization in medical images, and specifically to a method for automatic localization in a medical image based on deep learning. The method comprises the following steps: establishing a digital analysis model, collecting a large amount of medical image data, and performing integrity evaluation and screening on the collected medical image data; and on the basis of the result of screening of medical image information. According to the present invention, by means of the powerful feature extraction and generalization capabilities of a deep learning network, a specific area in a medical image can be automatically and accurately localized, avoiding the interference of errors and improving the accuracy of localization, by automatically localizing the specific area, the efficiency of medical diagnosis is improved, the deep learning network has strong adaptability and can process different types of medical images, unknown image information can also be fed back, and feedback data is collected and inputted into the model for system update, enabling the present invention to have a learning ability, thereby achieving both universality and flexibility.

Description

Automatic positioning method of medical images based on deep learning Technical Field

[0001] The present invention relates to the technical field of automatic positioning of medical images, and in particular to a method for automatic positioning of medical images based on deep learning. Background Art

[0002] In medical diagnosis, locating specific disease or abnormal areas is crucial. However, traditional manual localization methods are subjective, time-consuming, and error-prone. With the rapid development of deep learning, deep neural networks have facilitated automatic localization in medical image processing. However, when deep neural networks analyze images, they are unable to locate new, unknown image information in a timely manner. This results in the subsequent automatic localization of unknown images, which cannot be recognized due to data unrecognition, thus affecting the effectiveness of the method. Therefore, this paper proposes an automatic localization method for medical images based on deep learning.

[0003] Summary of the Invention

[0004] The purpose of the present invention is to provide a deep learning-based automatic positioning method for medical images to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, a method for automatic positioning of medical images based on deep learning is provided, which includes the following steps:

[0006] S1. Establish a digital analysis model, collect a large amount of medical image data, and conduct a complete evaluation and screening of the collected medical image data;

[0007] S2, based on the screening results of the medical image information in S1, pre-processing the retained medical image information and transmitting it to the digital analysis model for classification;

[0008] S3, collecting and evaluating the medical images to be tested, uploading the evaluation results to the digital analysis model established in S1, and performing feature extraction on the uploaded medical images using the digital analysis model;

[0009] S4, analyzing the medical image feature information extracted in S3, and selecting corresponding classification data for evaluation based on the analysis results, thereby obtaining a specific area of ​​the medical image;

[0010] S5, uploading the specific area information obtained in S3 to the cloud for evaluation, judging the accuracy of the specific area based on the evaluation results, and then performing secondary automatic positioning;

[0011] S6: Evaluate the specific area information of the image obtained based on the secondary automatic positioning of S5, save the target image and the positioning process according to the evaluation results, thereby obtaining accurate automatic positioning information of the medical image, and upload it to the digital model established in S1 for data update.

[0012] As a further improvement of this technical solution, the steps of transmitting S1 to the digital analysis model and classifying it are as follows:

[0013] S1.1. Establish a digital model for rapid image analysis corresponding to medical images;

[0014] S1.2. Collect medical image information from the existing network, analyze the collected medical image information based on integrity, and determine whether to retain the medical image information based on the analysis results.

[0015] As a further improvement of this technical solution, the steps of transmitting S2 to the digital analysis model and classifying it are as follows:

[0016] S2.1. Upload the medical image information retained in S1.2 to the image rapid analysis digital model established in S1.1 and perform preprocessing;

[0017] S2.2. Perform type analysis on the image data after preprocessing in S2.1, and classify the medical image information according to different types.

[0018] As a further improvement of this technical solution, the steps of collecting and evaluating the S3 medical image are as follows:

[0019] S3.1. Collect image information of the object to be detected;

[0020] S3.2. Evaluate the integrity of the object image information collected in S3.1 and conduct secondary object acquisition based on the evaluation results.

[0021] As a further improvement of this technical solution, the steps of S3 extracting features from the uploaded medical image using a digital analysis model are as follows:

[0022] S3.3, upload the latest object image information collected in S3.2 to the digital model for rapid image analysis established in S1.1;

[0023] S3.4. Analyze the latest object image information collected in S3.2 and perform feature extraction.

[0024] As a further improvement of the present technical solution, the step of obtaining the specific area of ​​the medical image in S4 is as follows:

[0025] S4.1. Analyze the characteristic data extracted in S3.2 and, based on the analysis results, search for classified data packets with corresponding properties within the digital analysis model;

[0026] S4.2. Combine the internal data of the corresponding data packet found in S4.1 and the feature data extracted in S3.2 to identify each other, thereby obtaining specific area information of the object image information in S3.1.

[0027] As a further improvement to this technical solution, the steps of determining the accuracy of the specific area based on the S5 evaluation result are as follows:

[0028] S5.1. Obtain specific region information of the object image information according to S4.2 and evaluate it in combination with known medical image information;

[0029] S5.2. Based on the evaluation results of S5.1, conduct a secondary analysis of the object image information in combination with the digital analysis model to perform specific area analysis.

[0030] As a further improvement of this technical solution, the steps of S6 in saving the target image and positioning process according to the evaluation results are as follows:

[0031] S6.1. Perform a judgment based on the latest specific area data obtained in S5.2, thereby determining the object image positioning data information based on the judgment result;

[0032] S6.2. Make a judgment based on the latest specific area data obtained in S5.2. If valid positioning information data cannot be obtained, save its image information and upload it to the cloud for feedback.

[0033] As a further improvement of this technical solution, the steps of uploading the data to the digital model in S6 for updating are as follows:

[0034] S6.3. Label specific areas of the object image information based on the feedback data obtained in S6.2;

[0035] S6.4. Combine the latest object image information collected in S3.2, the specific area information of the object image information obtained in S4.2, and the annotation information in S6.3, and upload the combined data to the image rapid analysis digital model in S1.1, thereby completing the data update.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] In this method for automatic positioning of medical images based on deep learning, by utilizing the powerful feature extraction and generalization capabilities of the deep learning network, it can automatically and accurately locate specific areas from medical images, avoid the interference of errors, and improve the accuracy of positioning. At the same time, by automatically positioning specific areas, the efficiency of medical diagnosis is improved. The deep learning network has strong adaptability and can process different types of medical images. By feeding back unknown image information and collecting feedback data and transmitting it into the model for system updating, the present invention has learning ability and realizes the combination of versatility and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] FIG1 is a flowchart of the overall process of the deep learning-based automatic positioning method for medical images of the present invention;

[0039] FIG2 is a flowchart of the complete evaluation and screening of collected medical image data according to the present invention;

[0040] FIG3 is a flow chart of the present invention for transmitting to a digital analysis model and performing classification;

[0041] FIG4 is a flowchart of the present invention for extracting features from uploaded medical images using a digital analysis model;

[0042] FIG5 is a flowchart of the present invention for obtaining a specific area of ​​the medical image;

[0043] FIG6 is a flow chart of the secondary automatic positioning process of the present invention;

[0044] FIG7 is a flowchart of data updating of the digital model of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0046] Example 1

[0047] Referring to FIG. 1 to FIG. 7 , the present embodiment aims to provide a method for automatic positioning of medical images based on deep learning, including the following steps:

[0048] S1. Establish a digital analysis model, collect a large amount of medical image data, and conduct a complete evaluation and screening of the collected medical image data;

[0049] The steps of transferring S1 to the digital analysis model and classifying it are as follows:

[0050] S1.1. Establish a digital model for rapid image analysis corresponding to medical images; use convolutional neural networks (CNNs) or deep learning models for image classification and segmentation. Through supervised learning, the model is trained on large-scale medical image data to automatically detect abnormalities or diagnose suspicious areas. Deep learning methods can adjust the depth and complexity of the network based on the complexity of the data and the number of samples, improving the accuracy of the algorithm, automating the entire image analysis process, and enhancing the precision and efficiency of diagnosis and treatment.

[0051] S1.2. Collect medical image information from the existing network and collect a large number of medical image data sets, including X-ray images, MRI images, and CT images. Analyze the collected medical image information based on integrity. The formula for medical image data integrity analysis is:

[0052] Perform quality assessment on medical images, comparing factors such as image detail, resolution, and noise to evaluate image integrity and usability.

[0053] Through image segmentation and measurement, information such as image morphology, size, and attributes can be extracted, the data quality of images of different bands and resolutions can be compared, and abnormal images can be marked.

[0054] According to the characteristics of the target medical images, the integrity standard and sampling strategy of the dataset are determined to maintain the integrity and reliability of the dataset.

[0055] Image processing and machine learning techniques are used to automatically classify and segment medical images to achieve complete analysis of large-scale medical image datasets.

[0056] Perform statistical analysis and visualization on the dataset, generate charts and reports to quickly evaluate the integrity and usability of the dataset, and provide valuable reference opinions to determine whether to retain medical image information based on the analysis results.

[0057] S2. Based on the screening results of the medical image information in S1, the retained medical image information is pre-processed and then transmitted to the digital analysis model for classification.

[0058] The steps of transferring S2 to the digital analysis model and classifying it are as follows:

[0059] S2.1. Upload the medical image information retained in S1.2 to the digital model for rapid image analysis established in S1.1 and perform preprocessing. Preprocess the collected medical image dataset: First, use a Gaussian filter to denoise the image, then apply a convolution kernel for smoothing. Finally, use histogram equalization to enhance image contrast. The Gaussian filter is a linear smoothing filter that removes noise by weighted averaging the image. After Gaussian filtering, smoothing can be performed by applying a convolution kernel. This is achieved by weighted averaging the area surrounding the pixel, producing a smoother image. Finally, histogram equalization is a technique for enhancing image contrast. It works by transforming the pixel values ​​of an image so that the distribution of pixel values ​​is as close to uniform as possible. This transformation can be achieved by calculating the cumulative distribution function (CDF) of the pixel values ​​in the histogram, and then using the pixel CDF value to remap the pixel values. Therefore, if you want to denoise, smooth, and enhance the contrast of an image, you can follow these steps:

[0060] Use Gaussian filter to denoise the image to obtain a clearer image by reducing noise.

[0061] Apply a convolution kernel for smoothing to further reduce sharp edges in the image, producing a smoother image.

[0062] Use histogram equalization to enhance the contrast of the image, making the image features more prominent and facilitating subsequent processing and analysis;

[0063] S2.2. Based on the image data pre-processed in S2.1, perform type analysis and classify the medical image information according to different types. According to the image properties, classify them into MRI images, X-ray images, and CT images to form MRI image data packets, X-ray image data packets, and CT image data packets;

[0064] S3, collecting and evaluating the medical images to be tested, uploading the evaluation results to the digital analysis model established in S1, and performing feature extraction on the uploaded medical images using the digital analysis model;

[0065] The steps for collecting and evaluating the S3 medical images are as follows:

[0066] S3.1. Collect image information of the object to be detected; extract the target object from the medical image, generally using an object detection algorithm. By identifying and locating the object of interest, such as a tumor or organ, from the image, the target is automatically located. The object detection algorithm uses the Faster R-CNN formula, which can be expressed in mathematical notation as follows:

[0067] First, use the pre-trained convolutional neural network to extract image features and obtain the feature map F;

[0068] Use RPN on the feature map F to obtain the candidate region;

[0069] For each candidate region, a fully connected layer is used to classify whether it contains an object, and regression is performed to obtain the coordinate information of the object;

[0070] Finally, for an image, the object coordinates and category information obtained will form an object detection output for object classification and precise positioning.

[0071] S3.2, conduct integrity assessment on the object image information collected in S3.1, and determine whether to re-collect the object based on the assessment results. Check whether the collected image has shadow parts, and if there are shadow parts, re-collect the image. If there are shadow parts, upload the data and feedback, and adjust the shadow parts of the object.

[0072] The steps for S3 to extract features from uploaded medical images using a digital analysis model are as follows:

[0073] S3.3, upload the latest object image information collected in S3.2 to the digital model for rapid image analysis established in S1.1;

[0074] S3.4. Analyze the latest object image information collected in S3.2 and perform feature extraction. Design a deep learning network structure based on the residual neural network (ResNet), including multiple convolutional layers, pooling layers, and fully connected layers. The network structure design takes into account the complexity and characteristics of medical image processing tasks. The basic formula of the deep learning network structure based on the residual neural network is as follows: For the input feature map x, it passes through a series of convolution operations and activation functions and then passes through multiple residual blocks, including:

[0075] The original input is extracted, representing the low-level features in the input: identity = x

[0076] Perform a series of convolution and activation operations to extract high-order features: y = F(x, W)

[0077] Construct a residual block based on high-order features: z = F(x, W) + identity, where z represents the predicted residual features

[0078] Add the residual feature z to the original input x to generate the final output feature: output = ReLU (z) + x, where W represents the weight parameter to be learned, F represents the combination of convolution and activation functions, and ReLU represents the rectified linear unit activation function. Based on the above formula, a deep neural network structure containing multiple residual blocks can be constructed. Each residual block repeats the above four basic steps, and multiple residual blocks are connected together through the above method to gradually generate more complex and abstract high-level features for tasks such as classification or regression. During the training process, the backpropagation algorithm is used to update the weight parameters of each residual block, and finally a high-precision deep learning model is trained, so that the network can effectively extract features from medical images;

[0079] S4, analyzing the medical image feature information extracted in S3, and selecting corresponding classification data for evaluation based on the analysis results, thereby obtaining a specific area of ​​the medical image;

[0080] The step of obtaining the specific area of ​​the medical image in S4 is as follows:

[0081] S4.1. Analyze the feature data extracted in S3.2 and, based on the analysis results, search for a classification data packet with corresponding properties within the digital analysis model; classify the image information into one of the following data packets based on the collected feature data: an MRI image data packet, an X-ray image data packet, and a CT image data packet;

[0082] S4.2, combine the internal data of the corresponding data packet found in S4.1 and the feature data extracted in S3.2 to identify the specific area information of the object image information in S3.1. The image rapid analysis digital model outputs the image based on the deep learning network structure of the residual neural network, locates the specific area in the medical image, and outputs the positioning result.

[0083] S5, uploading the specific area information obtained in S3 to the cloud for evaluation, judging the accuracy of the specific area based on the evaluation results, and then performing secondary automatic positioning;

[0084] The steps for determining the accuracy of the specific area based on the S5 evaluation results are as follows:

[0085] S5.1. Evaluate the specific region of the object image information obtained in accordance with S4.2 in combination with known medical image information; compare the specific region with the known medical image information within the digital analysis model. The formula for comparing the acquired medical image with the known medical image information includes the following steps:

[0086] Define the characteristics and classification factors of medical images, such as the comparison of lung cancer CT images.

[0087] Preprocess the collected medical images and known medical images, such as image denoising and image alignment.

[0088] Extract feature information of collected medical images and known medical images, such as morphology, size, texture, density, etc.

[0089] The collected medical images and known medical images are classified and discriminated, the feature data is input into the trained model, and the classification results of the medical images are given based on the model.

[0090] Analyze the results of the comparison, such as through accuracy calculation, error analysis and other methods, to evaluate the results of the comparison tasks.

[0091] Make medical judgments and treatments based on the comparison results, such as detection, diagnosis, and lesion localization based on lung cancer image comparison. If the comparison result is unknown, immediately perform a secondary image acquisition test to avoid image analysis errors.

[0092] S5.2. Based on the evaluation results of S5.1, conduct a secondary analysis of the object image information in combination with the digital analysis model to perform specific area analysis.

[0093] S6: Evaluate the specific area information of the image obtained based on the secondary automatic positioning of S5, save the target image and the positioning process according to the evaluation results, thereby obtaining accurate automatic positioning information of the medical image, and upload it to the digital model established in S1 for data update.

[0094] The steps of S6 for saving the target image and positioning process according to the evaluation results are as follows:

[0095] S6.1. Perform a judgment based on the latest specific area data obtained in S5.2, thereby determining the object image positioning data information based on the judgment result;

[0096] S6.2. Make a judgment based on the latest specific area data obtained in S5.2. If valid positioning information data cannot be obtained, save and upload the image information to the cloud for feedback. When the collected medical image information does not match the known image information, it can be defined as unknown image information and saved and uploaded to the cloud for subsequent analysis and comparison. The method for saving and uploading unknown image information is as follows:

[0097] First, create a folder in a local or cloud storage space specifically for storing unknown image information to facilitate storage and management.

[0098] Import unknown medical image information into the folder and store it in an image format (such as JPG, PNG, DICOM, etc.) or compress it into a ZIP file to reduce the file size.

[0099] For known classification features of medical image information, you can customize attribute labels such as patient name, disease type, collection instrument, hospital name, etc. Data standardization and normalization will help improve the efficiency of future medical data processing and the quality of data analysis.

[0100] An index file and a metadata file are created in the saved folder. The index file can customize the classification labels of unknown image information to facilitate subsequent search and retrieval, while the metadata file can describe and identify each medical image information stored in the folder.

[0101] Uploading unknown medical image information to a cloud storage platform allows for easy access and use whenever needed, while also increasing information backup and security. In summary, for situations where medical image information mismatches occur, standardization, normalization, and cloud storage can be used to preserve and manage unknown image information, facilitating subsequent processing and analysis. This information can also be fed back to users, who can interpret the unknown image information to improve the accuracy of subsequent image analysis.

[0102] The steps of uploading the S6 to the digital model for data update are as follows:

[0103] S6.3. Label specific areas of the object image information based on the feedback data obtained in S6.2;

[0104] S6.4. Combine the latest object image information collected in S3.2, the specific area information of the object image information obtained in S4.2, and the annotation information in S6.3, and upload the combined data to the image rapid analysis digital model in S1.1, thereby completing the data update.

[0105] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A medical image automatic positioning method based on deep learning, characterized by: The following steps are involved: S1. Establish a digital analysis model, collect a large amount of medical image data, and conduct a complete evaluation and screening of the collected medical image data; S2, based on the screening result of the medical image information in S1, pre-processing the retained medical image information and transmitting it to the digital analysis model for classification; S3, collecting and evaluating the medical images to be tested, uploading them to the digital analysis model established in S1 according to the evaluation results, and extracting features of the uploaded medical images through the digital analysis model; S4, analyzing the medical image feature information extracted in S3, and selecting corresponding classification data for the medical image for evaluation according to the analysis result, so as to obtain a specific area of ​​the medical image; S5, uploading the specific area information obtained by S3 to the cloud for evaluation, judging the accuracy of the specific area according to the evaluation result, and performing secondary automatic positioning; S6, based on the image specific area information obtained by the secondary automatic positioning of S5, the target image and the positioning process are saved according to the evaluation results, so as to obtain the accurate automatic positioning information of the medical image, and upload it to the digital model established in S1 for data update.

2. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of transferring S1 to the digital analysis model and classifying it are as follows: S1.

1. Establish a digital model for rapid image analysis corresponding to medical images; S1.

2. Collect medical image information from the existing network, analyze the collected medical image information based on integrity, and determine whether to retain the medical image information based on the analysis results.

3. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of transferring S2 to the digital analysis model and classifying it are as follows: S2.1, upload the medical image information retained in S1.2 to the image rapid analysis digital model established in S1.1 and perform preprocessing; S2.

2. Perform type analysis on the image data preprocessed in S2.1, and classify the medical image information according to different types.

4. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of collecting and evaluating the S3 medical images are as follows: S3.

1. Collect image information of the object to be detected; S3.

2. Evaluate the integrity of the object image information collected in S3.1, and perform secondary object collection based on the evaluation results.

5. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of S3 extracting features from uploaded medical images through a digital analysis model are as follows: S3.3, upload the latest object image information collected in S3.2 to S1.1 to establish a digital model for rapid image analysis; S3.4, analyze the latest object image information collected in S3.2 and perform feature extraction.

6. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The step of S4 acquiring the specific area of ​​the medical image is as follows: S4.1, analyzing the characteristic data extracted in S3.2, and searching for the classification data packets of corresponding properties in the digital analysis model based on the analysis results; S4.2, combining and identifying the internal data of the corresponding data packet found in S4.1 and the feature data extracted in S3.2, so as to obtain the specific area information of the object image information in S3.

1.

7. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of determining the accuracy of the specific area based on the S5 evaluation result are as follows: S5.1, obtaining specific region information of the object image information according to S4.2 and evaluating it in combination with known medical image information; S5.

2. Based on the evaluation results of S5.1, conduct a second analysis of the object image information in combination with the digital analysis model to perform specific area analysis.

8. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of S6 for saving the target image and positioning process according to the evaluation results are as follows: S6.1, making a judgment based on the latest specific area data obtained in S5.2, thereby determining the object image positioning data information according to the judgment result; S6.

2. Make a judgment based on the latest specific area data obtained in S5.

2. If no valid positioning information data can be obtained, Its image information is saved and uploaded to the cloud for feedback.

9. The method for automatic positioning of medical images based on deep learning according to claim 1, characterized in that: The steps of uploading S6 to the digital model for data update are as follows: S6.3, labeling specific areas of the object image information based on the feedback data obtained in S6.2; S6.4, combining the latest object image information collected by S3.2, the specific area information of the object image information obtained by S4.2, and the annotation information of S6.3, and uploading the combined data to the image rapid analysis digital model of S1.1, thereby completing the data update.

Images