Search

CN-121999491-A - Automatic identification and classification detection method for clinical urine formed components

CN121999491ACN 121999491 ACN121999491 ACN 121999491ACN-121999491-A

Abstract

The invention discloses an automatic identification and classification detection method for clinical urine components, and relates to the fields of clinical examination technology and artificial intelligent detection. The method comprises the steps of establishing a urine existing formation fractional data set containing time sequences, marking categories and positions, dividing a training set and a testing set, constructing a composite model integrating a transducer detection model, a Kalman filtering tracking model and a multi-feature recognition model, achieving target tracking and comparison through context correlation matching, collecting videos of urine samples to be tested, extracting single-frame images, positioning formed components after model detection, prediction and matching, comparing the types and the numbers of the data sets, and outputting results. The invention combines deep learning and dynamic tracking technology, balances detection speed and precision, reduces interference influence, has omission ratio less than or equal to 2 percent, accuracy rate more than or equal to 96 percent, accords with WS/T229-2024 standard, can adapt to various clinical urine samples, provides reliable basis for diagnosing urinary system diseases, and has important clinical application value.

Inventors

  • LI SHAOSHI
  • WU XUE

Assignees

  • 镇雄县人民医院

Dates

Publication Date
20260508
Application Date
20260228

Claims (5)

  1. 1. The automatic identification and classification detection method for the formed components of the clinical urine is characterized by comprising the following steps of: S1, constructing a tangible component data set: Extracting single-frame image data from the urine video according to time sequence, marking position coordinates of urine tangible components in each single-frame image and category information to form a formed fractional dataset, dividing the dataset into a training set and a testing set according to a ratio of 7:3, wherein the categories comprise red blood cells, white blood cells, epithelial cells, tubular cells, crystals and pathogens; S2, constructing and training a detection classification model: constructing a composite model integrating detection, tracking, prediction, context association matching and recognition functions, and specifically comprising the following steps: S21, constructing a model, wherein the composite model comprises a transducer detection module, a Kalman filtering tracking module and a multi-feature recognition module, wherein the transducer detection module consists of an encoder, a decoder and a multi-layer neural network mapping layer, and the encoder comprises at least two multi-head self-attention modules and is used for extracting image block features; S22, model training, namely inputting a training set into a model, acquiring a target detection frame coordinate and a change rate through a transducer detection module, after initializing a tracking module, alternately executing detection and state prediction on subsequent frame images, adopting Manhattan distance and twin network similarity calculation and a Hungary algorithm to realize context association matching, dividing matching targets, possibly losing targets and losing targets, updating the model, grouping by an identification module to infer target types, verifying and optimizing model parameters through a test set until the accuracy is more than or equal to 96%; S3, detecting a sample to be detected: repeating the video acquisition and single frame extraction of step S1 for the urine sample to be tested, inputting the trained model, detecting, tracking, matching, positioning all the formed components, numbering, comparing with the data set, identifying the types and the number, and outputting the detection result.
  2. 2. The method for automatic identification and classification detection of formed clinical urine according to claim 1, wherein in step S21, the tracking module may be replaced by a tracking model based on SSD algorithm, FPN algorithm, YOLO series algorithm or Mask R-CNN algorithm, and the identification module optimizes the fusion texture feature and convolutional neural network feature by using a minimum redundancy maximum correlation method.
  3. 3. The automatic identification and classification detection method for the formed components of the clinical urine according to claim 1, wherein in the step S1, a microscopic imaging device adopts a 10X-40X objective lens, the video acquisition frame rate is 25-30 frames/second, the single-frame image resolution is not lower than 1024X 768, and the labeling precision error is less than or equal to 2 pixels.
  4. 4. The method for automatically identifying and classifying the presence of a component in clinical urine according to claim 1, wherein in step S22, the weight matrix of the context-dependent matching is obtained by weighting the similarity and the manhattan distance, the weighting coefficients are respectively 0.6 and 0.4, and the missing target determination condition is a continuous 3-frame no-match result.
  5. 5. The method for automatically identifying and classifying the formed components of the clinical urine according to claim 1, wherein in the step S3, the output of the detection result includes the type, the number, the morphological characteristics and the confidence level of the formed components, and the target with the confidence level lower than 85% is marked as the suspicious item and prompts the manual review.

Description

Automatic identification and classification detection method for clinical urine formed components Technical Field The invention relates to the field of clinical examination technology and artificial intelligence detection, in particular to an automatic identification and classification detection method for clinical urine formed components. Background Urine component detection is an important means for diagnosing urinary system diseases and monitoring curative effects, and the accuracy of results directly influences clinical diagnosis and treatment decisions. The traditional detection method relies on manual microscopic examination, after urine sediment is prepared through centrifugation, inspection staff observe and count under a microscope, the method is greatly influenced by personnel experience and subjective judgment, the efficiency is low, the time consumption for single sample detection is usually 5-8 minutes, and the detection requirement of hundreds of samples in three-level hospitals in daily life is difficult to meet. Meanwhile, the manual microscopic examination has insufficient recognition sensitivity to trace formed components (such as a small quantity of tube type red blood cells and special-shaped red blood cells), is easy to miss judgment due to visual fatigue, has the problem of poor consistency of results due to uneven qualification of inspectors in a basic laboratory, and cannot provide a unified reference basis for cross-hospital diagnosis and treatment. Existing automatic detection techniques are mainly divided into two types, namely flow cytometry and digital imaging techniques. Although the detection speed of flow cytometry is high, the time consumption of a single sample can be controlled within 1 minute, the particle size and the components can be distinguished only through fluorescent staining and scattered light signals, the distinguishing capability of formed components (such as squamous epithelial cells and small epithelial cells, transparent tubes and mucus wires) similar to the morphology is weak, misjudgment is easy to occur, morphological image evidence cannot be provided, and the requirements of clinic tracing and rechecking of abnormal samples are not met. The digital imaging technology realizes automatic classification based on an image recognition algorithm, but has obvious defects that the particle position in a urine sample is easy to change along with the liquid flow, so that detection omission and repeated detection are caused, uneven illumination, sample impurity interference and formed component long tail distribution characteristics (such as extremely low ratio of a tube type pathogen to a sample) are reduced, the recognition precision is reduced, the detection speed and the detection accuracy are difficult to balance by a single algorithm, the recognition capability of micro components is sacrificed for pursuing the speed by most equipment, or the detection time is prolonged for improving the precision, and the clinical complex sample scene cannot be completely adapted. In addition, the existing digital imaging equipment has poor adaptability to high-concentration proteinuria and hematuria samples, is easy to lose efficacy in feature extraction caused by particle stacking, and needs an automatic identification and classification method with high precision, high stability and clinical suitability. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an automatic identification and classification detection method for the formed components of the clinical urine, solves the problems of missed detection, repeated detection and insufficient precision, realizes the optimal balance of speed and precision, and meets the standardized requirements of clinical examination. In order to achieve the above purpose, the present invention provides the following technical solutions: the automatic identification and classification detection method for the formed components of the clinical urine comprises the following steps: S1, constructing a tangible component data set: Extracting single-frame image data from the urine video according to time sequence, marking position coordinates of urine tangible components in each single-frame image and category information to form a formed fractional dataset, dividing the dataset into a training set and a testing set according to a ratio of 7:3, wherein the categories comprise red blood cells, white blood cells, epithelial cells, tubular cells, crystals and pathogens; S2, constructing and training a detection classification model: constructing a composite model integrating detection, tracking, prediction, context association matching and recognition functions, and specifically comprising the following steps: S21, constructing a model, wherein the composite model comprises a transducer detection module, a Kalman filtering tracking module and a multi-feature recognition module, wherein the transducer detection m