CN-121983207-A - Sample detection control method and system

CN121983207ACN 121983207 ACN121983207 ACN 121983207ACN-121983207-A

Abstract

The invention belongs to the technical field of sample detection, and provides a sample detection control method and system, wherein the method comprises the steps of constructing a sample detection model based on historical sample detection data, identifying a misjudgment sample of the sample detection model based on the sample detection model and combining a manual rechecking result of the sample, and evaluating the sample misjudgment degree of the sample detection model; if the sample misjudgment rate is high, the misjudgment samples are classified according to the misjudgment type, and the cluster analysis results of the historical detection image samples are combined to identify the outlier samples in the misjudgment samples, the feature deviation interference feature judgment is carried out based on the outlier samples, an outlier sample feature library is constructed, the sample features are matched with the outlier sample feature library in the subsequent sample detection process based on the outlier sample feature library, and the sample misjudgment rate is reduced. According to the invention, by constructing the outlier sample feature library, a targeted support is provided for model retraining through feature matching in the follow-up process, and the recognition capability of the model to rare and special feature samples can be effectively improved.

Inventors

CHEN LI
LIANG YAN
DING RUI

Assignees

徐州市第一人民医院

Dates

Publication Date: 20260505
Application Date: 20260312

Claims (10)

1. A sample detection control method is characterized by comprising the following steps: based on historical sample detection data, a sample detection model is constructed, based on the sample detection model and combined with a manual rechecking result of a sample, a misjudgment sample of the sample detection model is identified, and the sample misjudgment degree of the sample detection model is evaluated; if the sample misjudgment rate is high, classifying the misjudgment samples according to the misjudgment type, and identifying outlier samples in the misjudgment samples by combining the historical detection image sample clustering analysis results; Based on the outlier sample, judging the characteristic deviation interference characteristic, constructing an outlier sample characteristic library, and based on the outlier sample characteristic library, matching the sample characteristic with the outlier sample characteristic library in the subsequent sample detection process, so that the sample misjudgment rate is reduced.
2. The method for controlling sample detection according to claim 1, wherein the process of constructing the sample detection model is as follows: Collecting historical detection image samples, preprocessing sample detection result data, namely qualified labeling samples and unqualified labeling samples judged by history, eliminating repeated samples, automatically extracting multi-scale features of indoor images by means of a YOLOv backstlane, adding a random overturn, illumination adjustment, gaussian blur and angle rotation data enhancement strategy, and carrying out noise reduction treatment on the fuzzy images to obtain a standardized historical image sample dataset; The method comprises the steps of dividing a standardized historical image sample dataset into a training set, a verification set and a test set according to a ratio of 7:2:1, selecting YOLOv models as sample detection models, taking characteristics of historical detection image samples as input, and taking qualified labeling samples and unqualified labeling samples corresponding to the historical detection image samples as output labels for training; The method comprises the steps of enabling the detection accuracy of a model on a verification set to reach 92% through the learning rate, a detection frame threshold value and a sample confidence coefficient threshold value of the verification set optimization model, adopting a test set for verification, enabling the model to output a prediction detection frame, a prediction label and a sample confidence coefficient of a test set sample, and completing model effect verification based on the result to obtain a sample detection model.
3. The method for controlling sample detection according to claim 2, wherein the step of identifying erroneous judgment samples of the sample detection model comprises the steps of: collecting indoor target detection image samples to be detected, performing preprocessing and feature extraction, inputting a sample detection model, and outputting a preliminary detection label corresponding to the target detection image samples and sample confidence, wherein the preliminary detection label comprises qualified labeling samples and unqualified labeling samples corresponding to the target detection image samples; Performing manual blind labeling rechecking on the target detection image sample, and determining a manual rechecking label of the target detection image sample by adopting layered sampling, wherein the manual rechecking label comprises a qualified labeling sample and an unqualified labeling sample corresponding to the target detection image sample; If the manual review label is different from the preliminary detection label output by the sample detection model, the corresponding target detection image sample is marked as a misjudgment sample.
4. The method for controlling sample detection according to claim 3, wherein the step of evaluating the degree of erroneous sample judgment of the sample detection model comprises the steps of: counting the number proportion of misjudgment samples in all target detection image samples to obtain the sample misjudgment rate of a sample detection model; If the sample misjudgment rate of the sample detection model is greater than or equal to the sample misjudgment rate threshold, the sample misjudgment rate of the sample detection model is high, otherwise, the sample misjudgment rate of the sample detection model is low.
5. The method for sample detection control according to claim 4, wherein the determining process of the result of the cluster analysis of the historical detection image samples is as follows: And carrying out cluster analysis on the historical detection image samples by using a DBSCAN clustering algorithm, and outputting a historical qualified sample cluster and a historical unqualified sample cluster.
6. The method for sample detection control according to claim 5, wherein the step of identifying the outlier sample in the erroneous judgment sample comprises the steps of: Taking the standardized historical image sample data set as reference distribution, extracting a misjudgment sample feature vector, analyzing and processing the misjudgment sample feature vector, and identifying an outlier candidate sample; and extracting the confidence coefficient of the outlier candidate sample, and if the sample confidence coefficient of the outlier candidate sample is smaller than a sample confidence coefficient threshold value, marking the corresponding outlier candidate sample as the outlier sample.
7. The method for sample detection control according to claim 6, wherein the identifying of the outlier candidate sample is: Calculating the Euclidean distance between the sample and each sample in the historical qualified sample cluster by using an Euclidean distance formula, and summing all the Euclidean distances between the sample and each sample in the historical qualified sample cluster to obtain an average Euclidean distance between the sample and each sample in the historical unqualified sample cluster; Setting a judging threshold value of a history qualified sample cluster and a judging threshold value of a history unqualified sample cluster, and if the average Euclidean distance of the history qualified sample cluster is more than or equal to the judging threshold value of the history qualified sample cluster and the average Euclidean distance of the history unqualified sample cluster is more than or equal to the judging threshold value of the history unqualified sample cluster, marking the misjudged sample as an outlier candidate sample.
8. The method for sample detection control according to claim 7, wherein the determining of the feature deviation interference feature based on the outlier sample comprises: extracting a multi-layer feature fusion result of YOLOv, taking account of pixel features and semantic features, obtaining feature vectors of all data in an outlier sample feature vector, a history qualified sample cluster and a history unqualified sample cluster, integrating the pixel feature vectors of all data in the history qualified sample cluster into a history qualified sample cluster pixel feature vector set, and respectively taking arithmetic average values of all dimensions of all vectors in the history qualified sample cluster pixel feature vector set to obtain a history qualified sample cluster center vector; And calculating the cosine similarity of the outlier sample feature vector and the central vector of the historical qualified sample cluster, judging that the outlier sample has the feature deviation interference feature if the cosine similarity is smaller than the cosine similarity mean value of the historical qualified sample cluster, and judging that the outlier sample does not have the feature deviation interference feature if the cosine similarity is greater than or equal to the cosine similarity mean value of the historical qualified sample cluster.
9. The method for sample detection control according to claim 8, wherein the process of constructing the outlier sample feature library comprises the steps of: The sample basic information layer is used for storing basic association information of an outlier sample and is used as a core index basis of a feature library, and the storage content of the basic information layer comprises a sample unique identifier ID, a sample type, a misjudgment type, acquisition metadata and a manual review remark, and the basic association information is stored by adopting a relational database and takes the sample unique identifier ID as a unique primary key; The core feature layer is used for storing core feature data directly related to outlier samples and feature deviation judgment, and the storage content comprises Conv5 layer pixel feature vectors, historical qualified sample cluster center vectors, cosine similarity and historical qualified sample cluster cosine similarity mean values of single outlier samples, wherein the pixel feature vectors and the cluster center vectors are high-dimensional vectors, and a vector database is used for storing the pixel feature vectors and the cluster center vectors; The judging result layer is used for storing characteristic deviation related judging information of the outlier sample, the storage content comprises characteristic deviation judging results, a relational database is used for storing the characteristic deviation judging results, and the correlation is established between the characteristic deviation judging result layer and the sample basic information layer and between the characteristic deviation judging result layer and the core characteristic layer through the unique sample identification ID.
10. A sample detection control system for performing the method of any of the preceding claims 1-9, the system comprising: the misjudgment evaluation module is used for constructing a sample detection model based on historical sample detection data, identifying a misjudgment sample of the sample detection model based on the sample detection model and combining a manual rechecking result of the sample, and evaluating the sample misjudgment degree of the sample detection model; if the sample misjudgment rate is high, classifying the misjudgment samples according to the misjudgment type, and identifying the outlier samples in the misjudgment samples by combining the historical detection image sample cluster analysis result; And the outlier sample feature library construction module is used for judging feature deviation interference features based on outlier samples, constructing an outlier sample feature library, and matching sample features with the outlier sample feature library in the subsequent sample detection process based on the outlier sample feature library so as to reduce the sample misjudgment rate.

Description

Sample detection control method and system Technical Field The invention belongs to the technical field of sample detection, and particularly relates to a sample detection control method and system. Background In the field of medical sample detection, such as pathological section analysis, medical image diagnosis, in-vitro diagnosis reagent interpretation and other scenes, the accuracy and reliability of a detection result are directly related to disease diagnosis, treatment decision and patient safety. With the development of accurate medical treatment and intelligent medical treatment, clinical and laboratory requirements for the generalization capability, robustness and misjudgment control capability of an automatic detection system are unprecedented. At present, the medical sample detection technology faces multiple inherent challenges that firstly, medical image data has obvious heterogeneity, samples derived from different equipment, imaging protocols and acquisition conditions have quality and characteristic differences, the existing preprocessing flow lacks a targeted standardization and enhancement strategy, so that a model training basis is unstable, secondly, positive and negative samples in a medical scene often show serious unbalance, particularly rare pathological changes, problematic cases or early focus samples are rare, a general data enhancement method is difficult to effectively simulate subtle changes of morphology and texture, so that the model has weak recognition capability on long-tail distributed samples, thirdly, most detection models directly adopt general architecture and parameters, do not carry out refined adjustment on characteristics of high noise, low contrast, complex structure and the like of medical images, and fourthly, the basic performance is limited, and fourthly, the most critical point is that the existing system lacks a set of systematic misjudgment sample recognition, attribution and treatment mechanism. When a false positive occurs in the model, it is difficult to quickly distinguish whether the confusion is due to common features or is due to "outlier" features (such as special variant forms, rare artifacts, or very early lesions) that are not fully expressed in the training set. The lack of capability makes iterative optimization of the model lack of directionality, and can not effectively accumulate and utilize the extremely valuable error cases, so that when complex, fuzzy or rare medical images are handled, the generalization of the model is insufficient, the misjudgment rate is high, the dual-severity requirements of clinical diagnosis on high sensitivity and high specificity are difficult to meet, and meanwhile, risks are brought to laboratory quality control and automatic auditing. Therefore, the invention provides a sample detection control method and a system. Disclosure of Invention In order to overcome the deficiencies of the prior art, at least one technical problem presented in the background art is solved. In a first aspect, the present invention provides a sample detection control method, including: based on historical sample detection data, a sample detection model is constructed, based on the sample detection model and combined with a manual rechecking result of a sample, a misjudgment sample of the sample detection model is identified, and the sample misjudgment degree of the sample detection model is evaluated; if the sample misjudgment rate is high, classifying the misjudgment samples according to the misjudgment type, and identifying outlier samples in the misjudgment samples by combining the historical detection image sample clustering analysis results; Based on the outlier sample, judging the characteristic deviation interference characteristic, constructing an outlier sample characteristic library, and based on the outlier sample characteristic library, matching the sample characteristic with the outlier sample characteristic library in the subsequent sample detection process, so that the sample misjudgment rate is reduced. The invention further provides a method for constructing a sample detection model, which comprises the following steps: Collecting historical detection image samples, preprocessing sample detection result data, namely qualified labeling samples and unqualified labeling samples judged by history, eliminating repeated samples, automatically extracting multi-scale features of indoor images by means of a YOLOv backstlane, adding a random overturn, illumination adjustment, gaussian blur and angle rotation data enhancement strategy, and carrying out noise reduction treatment on the fuzzy images to obtain a standardized historical image sample dataset; The method comprises the steps of dividing a standardized historical image sample dataset into a training set, a verification set and a test set according to a ratio of 7:2:1, selecting YOLOv models as sample detection models, taking characteristics of historical detection imag