CN-121236009-B - CT image identity identification method, model training method, medium and device

CN121236009BCN 121236009 BCN121236009 BCN 121236009BCN-121236009-B

Abstract

The application provides a CT image identity recognizing method, a model training method, a medium and equipment, which comprise the steps of obtaining at least one group of image pairs to be recognized, wherein each group of image pairs to be recognized comprises two CT images to be recognized, recognizing the identity of the at least one group of image pairs to be recognized according to a target recognition model, obtaining a recognition probability vector, wherein the recognition probability vector corresponds to the prediction confidence of at least one identity recognition result, obtaining the identity recognition result of the at least one group of image pairs based on the recognition probability vector, and obtaining target interpretation data of the at least one identity recognition result according to the target recognition model. The application can realize end-to-end CT image identity identification without constructing a reference library, strengthen the pixel level contrast capability of a high-specificity local anatomical structure, improve the identification accuracy and model interpretability under the condition of a small sample, and meet the technical requirements of judicial identification scenes on efficient, objective and credible identity identification.

Inventors

LIN JUNYI
Dou Runting
JIANG JIEQING
SHEN YIWEN
ZHOU TIANYU

Assignees

复旦大学

Dates

Publication Date: 20260512
Application Date: 20250916

Claims (7)

1. A method for identifying identity of a CT image, the method comprising: Acquiring at least one group of image pairs to be identified, wherein each group of image pairs to be identified comprises two CT images to be identified; Carrying out identity recognition on the at least one group of image pairs to be recognized according to a target recognition model, and obtaining a recognition probability vector, wherein the recognition probability vector corresponds to the prediction confidence of the identity recognition result of the at least one group of image pairs to be recognized; Acquiring identity recognition results of the at least one set of image pairs to be recognized based on the recognition probability vector, and Acquiring target interpretation data of the identity recognition result of the at least one group of image pairs to be recognized according to the target recognition model, wherein the acquiring the recognition probability vector comprises the following steps: Respectively carrying out convolutional coding on the at least one group of image pairs to be identified according to the encoder of the target identification model so as to correspondingly acquire at least one group of output characteristic pairs; acquiring at least one group of detail feature pairs transmitted by the encoder according to a decoder of the target recognition model, wherein each group of detail feature pairs is used for indicating spatial position information of the image pairs to be recognized of a corresponding group; decoding a corresponding set of the output feature pairs based on the at least one set of detail feature pairs according to the decoder to obtain at least one set of feature map pairs; Obtaining a distance metric for the at least one set of feature map pairs to convert the at least one set of feature map pairs to at least one classification input based on the distance metric correspondence; And classifying the at least one classification input according to the classification head of the target recognition model to acquire the recognition probability vector.
2. The CT image identity qualification method of claim 1, wherein classifying the at least one classification input according to a classification header of the object qualification model comprises: Convolving the at least one classification input according to a first convolution layer of the classification head to obtain at least one first classification characteristic diagram; After synchronous batch standardization processing and activation processing are carried out on the at least one first classification characteristic diagram, a second convolution layer is input, and at least one second classification characteristic diagram is obtained; And after synchronous batch standardization processing and activation processing are carried out on the at least one second classification characteristic diagram, inputting the at least one second classification characteristic diagram into a global pooling layer to obtain the identification probability vector.
3. The CT image identity recognizing method according to claim 2, wherein the target interpretation data includes at least first target interpretation data and second target interpretation data, the first target interpretation data being acquired based on the second classification feature map, the second target interpretation data being acquired based on the feature map pair, wherein, Calculating gradient information of the identification probability vector relative to the at least one second classification characteristic map to obtain at least one first target interpretation data; Gradient information of the identification probability vector relative to the at least one set of feature map pairs is calculated to obtain at least one pair of second target interpretation data.
4. The CT image identity verification method according to claim 1, wherein the training method of the object verification model comprises: Acquiring at least one group of first training image pairs, wherein each group of first training image pairs comprises two first CT training images; Carrying out identity recognition on the at least one group of first training image pairs according to an original recognition model to obtain a first training probability vector, wherein the first training probability vector corresponds to the prediction confidence of a first identity recognition training result of the at least one group of first training image pairs; and adjusting parameters of the original identification model according to the first training probability vector so as to acquire the target identification model.
5. The CT image identity verification method of claim 4, further comprising: acquiring first CT images of a plurality of identical parts shot by a plurality of individuals at different time points; screening candidate areas based on each first CT image, and extracting target areas based on the candidate areas to obtain a plurality of first CT training images; At least one set of first CT training image pairs is arbitrarily selected and each set of said first CT training image pairs is subjected to a rigid calibration and windowed adjustment as said first training image pairs.
6. An electronic device, comprising: One or more processors, and One or more memories having stored therein computer readable code which, when executed by the one or more processors, implements the method of any of claims 1 to 5.
7. A computer readable storage medium having instructions stored thereon, which when executed by a processor, cause the processor to perform the method of any of claims 1 to 5.

Description

CT image identity identification method, model training method, medium and device Technical Field The application belongs to the technical field of forensics, and particularly relates to a CT image identity identification method, a model training method, a medium and equipment. Background In forensic clinical practice, individual identity recognition techniques are of great value. Currently, intelligent work of identity identification is mainly served for cadaver source identification of forensic pathology. The technology mainly depends on stable and unique anatomical features in an individual body, such as sinus cavity morphology of skull, physiological variation of bones, surgical implants, healing traces of old fracture and the like, and three-dimensional reconstruction and feature point matching are carried out on medical images such as CT and the like through a computer-aided technology, so that whether an unknown remains and a target individual are the same person or not is judged. The method is particularly suitable for identifying remains with serious putrefaction, destruction or incomplete, and has become an important tool in the field of forensic human science and radiology intersection. However, in forensic clinical practice, some of the works need to identify identity for the purpose of medical image authenticity identification, for example, whether fraudulent activities such as "falsifying damage evidence by falsifying other images" exist needs to be screened, and at this time, whether the main body of medical images such as CT is the same person needs to be judged. Therefore, how to accurately and efficiently identify the identity of CT images is a technical problem to be solved by those skilled in the art. Disclosure of Invention The application provides a CT image identity identification method, a model training method, a medium and equipment, which are used for solving the technical problem of how to accurately and efficiently identify the CT image identity. In a first aspect, the present application provides a method for identifying identity of CT images, the method comprising: Acquiring at least one group of image pairs to be identified, wherein each group of image pairs to be identified comprises two CT images to be identified; Carrying out identity recognition on the at least one group of image pairs to be recognized according to a target recognition model, and obtaining a recognition probability vector, wherein the recognition probability vector corresponds to the prediction confidence of the identity recognition result of the at least one group of image pairs to be recognized; Acquiring identity recognition results of the at least one set of image pairs to be recognized based on the recognition probability vector, and Acquiring target interpretation data of identity recognition results of the at least one group of image pairs to be recognized according to the target recognition model In one implementation manner of the first aspect, acquiring the recognition probability vector includes: The acquiring of the identification probability vector comprises the following steps: Respectively carrying out convolutional coding on the at least one group of image pairs to be identified according to the encoder of the target identification model so as to correspondingly acquire at least one group of output characteristic pairs; acquiring at least one group of detail feature pairs transmitted by the encoder according to a decoder of the target recognition model, wherein each group of detail feature pairs is used for indicating spatial position information of the image pairs to be recognized of a corresponding group; decoding a corresponding set of the output feature pairs based on the at least one set of detail feature pairs according to the decoder to obtain at least one set of feature map pairs; Obtaining a distance metric for the at least one set of feature map pairs to convert the at least one set of feature map pairs to at least one classification input based on the distance metric correspondence; And classifying the at least one classification input according to the classification head of the target recognition model to acquire the recognition probability vector. In one implementation manner of the first aspect, classifying the at least one classification input according to the classification header of the target recognition model includes: Convolving the at least one classification input according to a first convolution layer of the classification head to obtain at least one first classification characteristic diagram; After synchronous batch standardization processing and activation processing are carried out on the at least one first classification characteristic diagram, a second convolution layer is input, and at least one second classification characteristic diagram is obtained; And after synchronous batch standardization processing and activation processing are carried out on the at le