CN-115908394-B - Fine fracture detection method and system based on target detection and graph annotation force mechanism

CN115908394BCN 115908394 BCN115908394 BCN 115908394BCN-115908394-B

Abstract

The invention discloses a method and a system for detecting a fine fracture based on target detection and a drawing force mechanism, wherein the method comprises the steps of acquiring a plurality of time sequence chest CT images, preprocessing the images, and constructing a training sample set by using the chest CT images marked with the fine fracture area; the method comprises the steps of training a fine fracture detection model by using a training sample set, extracting image features by using a convolutional neural network, extracting a pattern meaning force matrix by using a pattern convolution neural network, combining the image features with the pattern meaning force matrix to obtain final features, combining a YOLO target detection network based on the final features, outputting a fine fracture region prediction frame, inputting a chest CT image of a time sequence to be detected into the trained fine fracture detection model, outputting a fine fracture region and a fracture type, optimizing an identification result by using a target tracking algorithm, improving the accuracy of fine fracture region identification, and realizing accurate positioning of the fine fracture region in the chest CT image.

Inventors

LI CHUANPENG
FAN ZHAOLEI
ZHANG SONG
GUO KAIFENG
ZHANG MIN
LIU ZHAOKANG
MENG CHUAN
CHEN HONG

Assignees

图灵医道医疗器械科技(上海)有限公司

Dates

Publication Date: 20260512
Application Date: 20221229

Claims (7)

1.A tiny fracture detection method based on target detection and a drawing force mechanism is characterized by comprising the following steps: acquiring a plurality of time sequence chest CT images, preprocessing the images, and constructing a training sample set by taking the chest CT images marked with the tiny fracture areas as training samples; inputting a training sample image into a fine fracture detection model, extracting image features by using a convolutional neural network, extracting a graph meaning force matrix by using a graph convolution neural network, combining the image features and the graph meaning force matrix to obtain final features, combining a YOLO target detection network based on the final features, outputting a fine fracture region prediction frame, and training the fine fracture detection model according to the distance loss between the prediction frame and the target frame; Inputting the chest CT image to be tested into a training fine fracture detection model, and outputting a fine fracture region and fracture types; after inputting the sequential chest CT image into the fine fracture detection model, sequentially outputting corresponding target detection prediction frames, inputting the output result into the fine fracture detection optimization model, and performing post-processing reinforcement on the whole image sequence prediction result by adopting an NMS algorithm and a target tracking algorithm based on the z direction; An NMS algorithm and a target tracking algorithm based on the z-direction, comprising: mapping the predicted frame position with the highest confidence in the previous chest CT image and the predicted frame position with the highest confidence in the next chest CT image into the current chest CT image, taking the predicted frame position with the highest confidence in the current chest CT image as a center point, and comparing the position with the mapped position; If the z coordinate of the mapping position is within the preset value range and the IOU of the prediction frame of the mapping position and the prediction frame of the central point position is larger than the set threshold, deleting the prediction frame of the mapping position.
2. The method for detecting a small fracture based on a target detection and a mindset force mechanism according to claim 1, wherein the preprocessing comprises: Stacking a plurality of time sequence chest CT images to synthesize a multi-channel image; resampling is carried out according to the distance between pixels of the chest CT image, and normalization is completed.
3. The method for detecting a small fracture based on a target detection and a graph meaning force mechanism according to claim 1, wherein the combination of the image feature and the graph meaning force matrix to obtain a final feature comprises: Calculating a graph meaning force matrix extracted from the graph convolution neural network by using 1X 1 convolution, and changing a final nonlinear activation function into sigmoid to enable the value range of a calculation result to be 0-1; Multiplying the calculated graph meaning force matrix with the feature graph extracted by the convolutional neural network to obtain the final feature.
4. The method for detecting a small fracture based on a target detection and a force-on-attention mechanism as recited in claim 1, wherein the range of values of the z-coordinate is determined according to the inter-pixel distance of the chest CT image.
5. A fine fracture detection system based on target detection and a mindset force mechanism, comprising: the training sample construction module is used for acquiring a plurality of time sequence chest CT images, preprocessing the images, and constructing a training sample set by taking the chest CT images marked with the tiny fracture areas as training samples; Inputting a training sample image into a fine fracture detection model, extracting image features by using a convolutional neural network, extracting a graph meaning force matrix by using a graph convolution neural network, combining the image features with the graph meaning force matrix to obtain final features, combining a YOLO target detection network based on the final features, outputting a fine fracture region prediction frame, and training the fine fracture detection model according to the distance loss between the prediction frame and the target frame; The fine fracture detection module is used for inputting the chest CT image to be detected into the trained fine fracture detection model and outputting a fine fracture region and fracture types; The system also comprises a fine fracture detection optimization module, a fine fracture detection optimization module and a target tracking module, wherein the fine fracture detection optimization module is used for inputting an output result into the fine fracture detection optimization model, and performing post-processing reinforcement on the prediction result of the whole image sequence by adopting an NMS algorithm and a target tracking algorithm based on the z direction; An NMS algorithm and a target tracking algorithm based on the z-direction, comprising: mapping the predicted frame position with the highest confidence in the previous chest CT image and the predicted frame position with the highest confidence in the next chest CT image into the current chest CT image, taking the predicted frame position with the highest confidence in the current chest CT image as a center point, and comparing the position with the mapped position; If the z coordinate of the mapping position is within the preset value range and the IOU of the prediction frame of the mapping position and the prediction frame of the central point position is larger than the set threshold, deleting the prediction frame of the mapping position.
6. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of a method for fine fracture detection based on a target detection and a force-on-drawing mechanism as claimed in any one of claims 1 to 4.
7. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of a method of fine fracture detection based on a target detection and a force-on-icon mechanism as claimed in any one of claims 1 to 4.

Description

Fine fracture detection method and system based on target detection and graph annotation force mechanism Technical Field The invention relates to the technical field of medical image processing, in particular to a method and a system for detecting fine fracture based on target detection and a drawing force mechanism. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. When the chest of a human body is attacked by external pressure or violence, tiny fracture often occurs at the hit part, the fracture end is broken inwards, and meanwhile, the internal organs of the chest can be damaged, so that the chest fracture has higher morbidity and certain death risk. Among them, rib fracture is the most common injury in chest fine fracture, and is one of the main contents of medical identification and judicial identification. At present, computer Tomography (CT) is a main method for diagnosing fine fracture of chest, and a doctor usually locates the fine fracture of chest and the position of its complications according to the conventional CT scan image of chest, but because the chest part is large, including various bones such as ribs, collarbone, sternum, vertebrae, etc., the fine fracture is usually hidden and difficult to be found, and the CT scan image of chest generally includes several tens to several hundreds of different images, so locating fracture position in each CT image of chest is a mechanical and tedious task, and if there are a large number of patients, the workload of detecting fine fracture of chest will be huge. Meanwhile, the unavoidable asthenopia exists in consideration of human detection and identification, so that missed diagnosis and misdiagnosis occur at present, and the accuracy of detecting the chest tiny fracture is low. With the rapid development of deep learning technology in medical imaging, more studies are carried out on detection of lesions such as lung nodules, chest fractures, breast molybdenum targets and the like, and higher effects are exerted in practical application. However, the method for automatically identifying the fracture parts of the chest in the prior art is to detect the CT images of the chest one by one, neglecting the time sequence relation among the images, neglecting the position information of the image features in the images in the detection process, and the position information is particularly important for positioning the small target, so that the accuracy of positioning the fracture position is lower. Disclosure of Invention In order to solve the defects in the prior art, the invention provides a method and a system for detecting fine fracture based on a target detection and attention mechanism, which are used for identifying fine fracture areas in chest CT images based on chest CT images by adopting a target detection algorithm of the attention mechanism, and simultaneously, taking time sequence relations among the chest CT images into consideration, optimizing an identification result by adopting a target tracking algorithm, improving the accuracy of fine fracture area identification and solving the problem of inaccurate positioning of the fine fracture areas in the chest CT images in the prior art. In a first aspect, the present disclosure provides a method for detecting a small fracture based on a target detection and a deliberate force mechanism, comprising: acquiring a plurality of time sequence chest CT images, preprocessing the images, and constructing a training sample set by taking the chest CT images marked with the tiny fracture areas as training samples; inputting a training sample image into a fine fracture detection model, extracting image features by using a convolutional neural network, extracting a graph meaning force matrix by using a graph convolution neural network, combining the image features and the graph meaning force matrix to obtain final features, combining a YOLO target detection network based on the final features, outputting a fine fracture region prediction frame, and training the fine fracture detection model according to the distance loss between the prediction frame and the target frame; Inputting the time sequence chest CT image to be detected into a training-completed fine fracture detection model, and outputting a fine fracture region and fracture types. According to a further technical scheme, the pretreatment comprises the following steps: Stacking a plurality of time sequence chest CT images to synthesize a multi-channel image; resampling is carried out according to the distance between pixels of the chest CT image, and normalization is completed. According to a further technical scheme, the combination of the image features and the annotation force matrix to obtain final features comprises the following steps: Calculating a graph meaning force matrix extracted from the graph convolution neural network by using 1