CN-122025040-A - Automatic marking method and system for focus in medical image

CN122025040ACN 122025040 ACN122025040 ACN 122025040ACN-122025040-A

Abstract

The invention discloses a focus automatic marking method and system in medical images, which relate to the technical field of medical equipment and comprise the following steps of collecting medical images of different modality types, preprocessing, manually marking focus areas to obtain marked medical images, calculating characteristic parameters of the gray marked medical images to obtain classification intensity coefficients, dividing initial training data and dynamic training data, constructing focus recognition submodels for each modality type, training by utilizing the initial training data, carrying out forward propagation reasoning, reading focus prediction probability values, calculating pixel prediction offset degree based on the prediction probability values, adjusting learning rate, inputting the dynamic training data into the submodels, iteratively executing until the offset degree is lower than a preset threshold value, finally realizing automatic focus marking of the medical images to be marked, providing more accurate marking results, and enhancing recognition accuracy of fuzzy focuses.

Inventors

ZHAO ZHENFENG
WANG SHENWEN
ZHANG WEIFENG
WANG CAILAN

Assignees

河北深昊智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (8)

1. An automatic marking method for focus in medical image is characterized by comprising the following specific steps: s1, acquiring a plurality of medical images of different modality types, preprocessing, marking focus areas in the medical images by adopting a manual marking method to obtain marked medical images, and marking the preprocessed medical images as original medical images; s2, marking the graying marked medical image as a marked gray image, calculating and analyzing image characteristic parameters of the marked gray image to obtain a classification intensity coefficient for representing the identification difficulty degree of the focus in the image, and dividing the corresponding original medical image into initial training data and dynamic training data; S3, constructing a focus recognition sub-model aiming at each mode type, training the focus recognition sub-model by utilizing initial training data of the corresponding mode type, inputting part of dynamic training data into the focus recognition sub-model which is completed training to perform forward propagation reasoning, and reading focus prediction probability values of all pixel points of an original medical image in the reasoning process; S4, calculating the prediction offset degree of each pixel point based on the focus prediction probability value of the input partial dynamic training data, and adjusting the learning rate of the corresponding focus recognition sub-model based on the prediction offset degree; s5, inputting the residual original medical images in the dynamic training data into the corresponding focus recognition submodel, repeatedly executing the step S4 until the prediction deviation degree does not exceed a preset threshold value, completing training of the focus recognition submodel, preprocessing the medical images to be marked based on the mode type of the medical images to be marked, inputting the preprocessed medical images to be marked into the corresponding focus recognition submodel, and outputting the medical images completing focus marking.
2. The method for automatically marking a focus in a medical image according to claim 1, wherein the preprocessing specifically refers to image noise reduction preprocessing, and light and shadow noise is eliminated; The modality types of the medical images include X-ray images, CT images, magnetic resonance imaging images, tissue slice images, and endoscopic images.
3. The method for automatically marking the focus in the medical image according to claim 2, wherein the specific logic of marking the focus area in the medical image by adopting a manual marking method is that a medical image marking tool is used for marking the focus area in the medical image through a minimum circumscribed rectangle, the marked medical image is archived, a file format is selected for storing a marking result, and the marked medical image and a corresponding original medical image form a mapping relation for training a focus recognition submodel; The image characteristic parameters of the marked gray image comprise gray significant factors, focal region area occupation ratio and gradient strength of the edge of the focal region; Calculating the gradient intensity of the edge of the focus area according to the following formula by calculating the gradient intensity of each pixel point of the edge of the focus area in the marked gray image through a Sobel operator, taking the average gradient intensity of all the edge pixel points as the gradient intensity of the edge of the focus area, and calculating the gradient intensity of the edge of the focus area according to the following formula: In the formula, Is the gradient intensity of the edge of the focal region, Gradient intensity of the ith edge pixel point of the focus area, wherein i is index of the edge pixel point of the focus area, For the total number of edge pixels in the lesion area, 。
4. The method for automatically marking a focus in a medical image according to claim 3, wherein the logic based on which the classification intensity coefficient is calculated is that a gray scale significant factor of a marked gray scale image, a focus area occupation ratio and a gradient intensity of a focus area edge are calculated, wherein the gray scale significant factor is generated by marking a difference between a gray scale value in the focus area and a gray scale value in a non-focus area in the gray scale image, the classification intensity coefficient is comprehensively generated according to the gray scale significant factor and the focus area occupation ratio and the gradient intensity of the focus area edge, the complexity of focus identification in the image is characterized by the classification intensity coefficient, and the gray scale significant factor is calculated according to the formula: In the formula, As a significant factor of the gray scale, To label the average gray value inside the lesion area in the gray image, The average gray value of a non-focus area in the same marked gray image is obtained; The focal region area ratio specifically refers to the ratio of the number of pixel points in the focal region to the total number of pixel points in the labeling gray level image to which the focal region belongs; the specific formula for calculating the classification intensity coefficient is as follows: In the formula, In order to classify the intensity coefficient of the light, The values are normalized for the gray scale saliency factors, Is normalized for the focal region area ratio, The method specifically based on the normalization of the gray scale significant factor, the focal region area ratio and the gradient intensity of the focal region edge is that the normalization of each parameter is carried out by adopting a minimum-maximum normalization method.
5. The method for automatically marking a focus in a medical image according to claim 4, wherein the logic based on which the original medical image is divided into the initial training data and the dynamic training data according to the classification intensity coefficients is that the classification intensity coefficients of all the marked gray level images are calculated and compared with a preset classification threshold value, and the corresponding original medical image is divided into the initial training data and the dynamic training data according to the comparison result, specifically: If it is Dividing the original medical image into initial training data; If it is The original medical image is divided into dynamic training data, A preset classification threshold value; The focus recognition sub-model is built based on a deep learning U-Net model, the segmentation loss and the focus loss are selected as loss functions, adam is an optimizer, the focus recognition sub-model is input into an original medical image of a corresponding mode type, and the focus recognition sub-model is output into a marked medical image of a marked focus area; The logic based on which the corresponding focus recognition sub-model is specifically trained through the initial training data is that the original medical image in the initial training data is used as input, the marked medical image of the focus area is used as a label, and each focus recognition sub-model is trained; the logic specifically based on forward propagation reasoning in the focus recognition submodel with the training completed by inputting the dynamic training data is that the Dropout layer is kept in an activated state during forward propagation reasoning, and focus prediction probability values of all pixel points of the original medical image after each reasoning are read through a Monte Carlo Dropout sampling mode.
6. The method for automatically marking a focus in a medical image according to claim 5, wherein the logic based on the specific calculation of the prediction offset degree of each pixel based on the focus prediction probability value is that in all reasoning processes, the focus prediction probability average value of each pixel is calculated, the entropy value of the focus prediction probability of each pixel is calculated according to the focus prediction probability average value, the prediction offset degree of each pixel is characterized by the entropy value of the focus prediction probability of each pixel, and the formula based on the specific calculation of the entropy value of the focus prediction probability of each pixel is as follows: In the formula, Representing coordinates as The pixel focus prediction probability entropy value of (1), Representing coordinates as Wherein x is a pixel point abscissa variable and y is a pixel point ordinate variable; Wherein the calculation is The specific formula is as follows: In the formula, Representing coordinates as The focus prediction probability value of the pixel point after the t-th reasoning, wherein t is the index of the reasoning turn, , Is the total number of inferences.
7. The method for automatically marking focus in medical image according to claim 6, wherein the specific basis of the logic for adjusting the learning rate of the corresponding focus recognition sub-model based on the prediction deviation degree is that in dynamic training data, the original medical image of a single training batch number is acquired, the original medical image is input into the corresponding focus recognition sub-model based on a mode type for incremental training, the learning rate is adjusted according to focus prediction probability average entropy values of all original medical images in the single training batch number, wherein the focus prediction probability entropy values of the single Zhang Yuanshi medical image are characterized by the average value of focus prediction probability entropy values of all pixel points in the focus prediction probability entropy values; The specific formula based on which the learning rate is adjusted is as follows: In the formula, In order to adjust the learning rate after the adjustment, For the initial learning rate in the pre-training process, The average entropy value of the probability of lesion prediction for all the original medical images in a single training, For single training of focus prediction probability entropy reference values of original medical images, Is a proportionality constant; The logic on which the specific basis of inputting the residual original medical images in the dynamic training data into the corresponding focus recognition submodel is that the residual original medical images in the dynamic training data are divided according to the number of single training batches to obtain a plurality of groups of increment training detection data groups, the increment training detection data groups are sequentially input into the focus recognition submodel after completing the current increment training, the focus prediction probability entropy value of each pixel point of the images in the forward propagation reasoning calculation reasoning process is carried out again, if the focus prediction probability entropy value of the single Zhang Yuanshi medical images is larger than a preset threshold value, the input image is marked, when the marked input image reaches the number of single training batches, the marked input image is used as the input data of a new round of increment training, and the learning rate is adjusted until the focus prediction probability entropy value of all the original medical images in one round of training is smaller than the preset threshold value, and the increment training is completed.
8. An automatic marking system for a focus in a medical image, characterized in that the automatic marking system for a focus in a medical image is used for executing the automatic marking method for a focus in a medical image according to any one of claims 1 to 7, comprising: The image mode identification module is used for acquiring a plurality of medical images of different mode types and preprocessing the medical images, marking focus areas in the medical images by adopting a manual marking method to obtain marked medical images, and marking the preprocessed medical images as original medical images; The training data planning module is used for marking the graying marked medical image as a marked gray image, calculating and analyzing image characteristic parameters of the marked gray image to obtain a classification intensity coefficient for representing the identification difficulty degree of the focus in the image, and dividing the corresponding original medical image into initial training data and dynamic training data; The prediction probability presumption module is used for constructing a focus recognition sub-model aiming at each mode type, training the focus recognition sub-model by utilizing initial training data of the corresponding mode type, inputting part of dynamic training data into the focus recognition sub-model which is completed with training to carry out forward propagation reasoning, and reading focus prediction probability values of all pixel points of an original medical image in the reasoning process; The learning rate adjustment module is used for calculating the prediction offset degree of each pixel point based on the focus prediction probability value of the input partial dynamic training data and adjusting the learning rate of the corresponding focus recognition sub-model based on the prediction offset degree; the iterative optimization module is used for inputting the residual original medical images in the dynamic training data into the corresponding focus recognition sub-model, repeatedly executing the function of the learning rate adjustment module until the prediction deviation degree does not exceed a preset threshold value, completing the training of the focus recognition sub-model, preprocessing the medical images to be marked on the basis of the mode type of the medical images to be marked, inputting the preprocessed medical images to be marked into the corresponding focus recognition sub-model, and outputting the medical images completing focus marking.

Description

Automatic marking method and system for focus in medical image Technical Field The invention relates to the technical field of medical equipment, in particular to an automatic focus marking method and system in medical images. Background In recent years, with the rapid development of medical imaging technology, the number and complexity of medical images has rapidly increased. This provides more possibilities for early diagnosis and treatment of diseases, but at the same time presents challenges. The traditional medical image analysis method depends on manual observation and experience judgment, and is difficult to meet the requirements of rapidness, accuracy and high efficiency. Especially in the aspect of identification and marking of lesions such as tumors, manual identification is time-consuming and is easily affected by subjective factors, and diagnosis omission or misdiagnosis can be caused. Therefore, there is a need for an efficient and accurate automated lesion image recognition method to improve the quality and efficiency of medical image analysis. In this context, the application of deep learning techniques brings new opportunities for medical image analysis. The existing deep learning model can learn features from a large number of medical images and automatically classify and identify the features. However, how to extract effective features from complex medical images, accurately identify different organs and tissues, and select an appropriate model for prediction according to the identification result is still a problem to be solved. In addition, how to ensure accuracy of the recognition result in order to effectively mark the medical image during the recognition process also faces technical challenges. In the prior art, publication number CN110232383a discloses a focal image recognition method and a focal image recognition system based on a deep learning model, specifically, for an acquired medical image to be detected, by sequentially performing segmentation of a minimum organ tissue image, image recognition of a corresponding organ tissue, prediction of the deep learning model and marking on a prediction result graph, a doctor can be replaced to find a potential disease condition in the medical image, and recognition marking of the focal tissue is automatically performed on the medical image so as to remind the doctor to make further diagnosis, timely confirm the disease condition, further reduce the working strength of the doctor, timely confirm whether the disease condition occurs or not, but in the scheme, the image is preferentially recognized by the image, the medical image may be affected by various noises, such as equipment noise, environmental noise, and the like, which may cause image quality to be reduced, further affect feature extraction and organ recognition, and different acquisition angles and positions may cause the same organ to present different forms and features in the image, thereby causing reduction of recognition accuracy, in the scheme is mainly dependent on uniform training deep learning model to make prediction, and the accuracy of the dynamic model cannot be adjusted according to training data relative to the training model, thereby reducing the accuracy of the dynamic model recognition performance of the common image. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide a method and a system for automatically marking a focus in a medical image, which are used for solving the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: An automatic marking method for focus in medical image comprises the following steps: s1, acquiring a plurality of medical images of different modality types, preprocessing, marking focus areas in the medical images by adopting a manual marking method to obtain marked medical images, and marking the preprocessed medical images as original medical images; s2, marking the graying marked medical image as a marked gray image, calculating and analyzing image characteristic parameters of the marked gray image to obtain a classification intensity coefficient for representing the identification difficulty degree of the focus in the image, and dividing the corresponding original medical image into initial training data and dynamic training data; S3, constructing a focus recognition sub-model aiming at each mode type, training the focus recognition sub-model by utilizing initial training data of the corresponding mode type, inputting part of dynamic training data into the focus recognition sub-model which is completed training to perform forward propagation reasoning, and reading focus predic