CN-122023374-A - Two-stage lightweight network-based pleuroperitoneal cavity effusion detection and segmentation method

CN122023374ACN 122023374 ACN122023374 ACN 122023374ACN-122023374-A

Abstract

The invention discloses a two-stage lightweight network-based pleuroperitoneal cavity effusion detection and segmentation method, which belongs to the technical field of medical image analysis and comprises the steps of constructing a training data set containing a plurality of groups of pleuroperitoneal cavity effusion data, constructing a two-stage deep learning network model, training the model by utilizing the training data set, inputting a to-be-processed pleuroperitoneal cavity image into a trained SL-YOLO detection model for detection, and inputting an image area intercepted by a detection frame into a trained SLNet segmentation model for segmentation.

Inventors

YUAN HONGBIN
LI YONGHUA
CHEN WEI

Assignees

中国人民解放军海军军医大学

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (7)

1. The utility model provides a method for detecting and segmenting pleuroperitoneal cavity hydrops based on two-stage lightweight network, which is characterized by comprising the following steps: S1, acquiring a plurality of groups of pleuroperitoneal cavity ultrasonic images, and marking a detection frame and marking a segmentation mask for a effusion area in the images by a professional to form a training data set; S2, constructing a two-stage deep learning network model, wherein the first stage adopts an SL-YOLO detection model designed for small target detection to position a effusion region, and the second stage adopts a light-weight improved SLNet segmentation model to accurately segment the effusion region; The SLNet segmentation model is a lightweight improved model of a U-Net architecture, the SLNet segmentation model performs feature extraction by adopting a multi-core depth separable convolution module to replace standard convolution in an encoding path, a gating network is introduced between an encoder and a decoder to perform jump connection of adaptive feature screening, and a region perception decoding layer is introduced in a decoding path to fuse global and local context features, so that the segmentation precision is kept while the computational complexity is reduced; S3, training the SL-YOLO detection model and the SLNet segmentation model by using a training data set; S4, inputting the to-be-processed pleuroperitoneal cavity image into a trained SL-YOLO detection model, and outputting a detection frame of one or more effusion detection areas; s5, inputting the image area intercepted by the detection frame into a trained SLNet segmentation model, and outputting a corresponding effusion area segmentation mask.
2. The method for detecting and segmenting the pleuroperitoneal cavity effusion based on the two-stage lightweight network of claim 1, wherein the training data set constructed in the step S1 comprises pleuroperitoneal cavity ultrasonic images of patients with different sexes, ages and medical histories, and the conditions of normal images, mild effusion and severe effusion are covered.
3. The method for detecting and segmenting pleuroperitoneal cavity effusion based on the two-stage lightweight network according to claim 1, wherein a convolutional layer of a SLNet segmentation model coding path adopts multi-core depth separable convolution MSDWC for feature extraction, and is specifically expressed as follows: ; Wherein, the A depth separable convolution representing a convolution kernel size k, The representation Relu activates the function, A batch normalization operation is represented and, Representing a set of convolution kernel sizes, Representing an input image; the convolution layer MSDWCL of the coding path is formed by sequentially connecting one or more multi-core depth separable convolutions MSDWC, and a specific calculation formula is as follows: ; Wherein, the A standard convolution operation is represented and, A combination of batch normalization operations and Relu activation functions is shown.
4. The method for detecting and segmenting pleuroperitoneal cavity effusion based on the two-stage lightweight network according to claim 3, wherein the specific operation flow of the region-aware decoding layer of the SLNet segmentation model is as follows: a. global features are extracted through global max-pooling GMP operation, and local features are extracted through local max-pooling LMP operation, specifically expressed as: ; ; Wherein, the And Representing the global feature and the local feature respectively, Representing a global maximization of the pool, Representing a local maximum pooling of the data, An expansion operation is represented by a sequence of expansion operations, Representing the input characteristics of the region sensing decoding layer after being processed by the gating network; b. adding the global features and the local features element by element, calculating to obtain a sensing weight through a Sigmoid activation function, and carrying out element by element product with the input features to obtain the output of the enhanced spatial region sensing module The method is specifically expressed as follows: ; Wherein, the Representing the per-element product, The representation of the sum from element to element, Representing a Sigmoid activation function; c. and decoding by a region sensing decoder to obtain final output, and sending the final output to a segmentation head for outputting the hydrops region segmentation mask.
5. The method for detecting and segmenting pleuroperitoneal cavity effusion based on two-stage lightweight network according to claim 4, wherein a gating network of SLNet segmentation models receives gating signals And convolutional layer features Processing and calculating gating attention weight through grouping convolution, and outputting gating characteristics, wherein the gating characteristics are specifically expressed as follows: ; ; ; Wherein, the A convolution of the packets is represented, The characteristics of the fusion are represented and, Representing the weight of the gated attention, Representing a gating feature.
6. The method for detecting and segmenting pleuroperitoneal cavity effusion based on the two-stage lightweight network according to claim 1, wherein the loss functions trained by the SL-YOLO detection model in S3 comprise position loss, confidence loss and category loss, specifically expressed as: ; Wherein, the The weight of the coordinate loss is represented, The weight representing the loss of confidence, The weight of the class loss is represented, Is a tag of a true class of tags, Is a label of the predictive category and, And Respectively whether the grid contains an object or not, And The confidence levels of reality and prediction respectively, And The actual and predicted bounding box parameters, respectively.
7. The method for detecting and segmenting the pleuroperitoneal cavity effusion based on the two-stage lightweight network according to claim 1, wherein the Loss function trained by the SLNet detection model in S3 comprises a binary cross entropy BCE Loss function, a DICE Loss function and a Focal Loss function, which are specifically expressed as: ; ; ; ; Wherein, the Representing a real label of the tag, The output result of the model is represented, Representing the number of groups of model predicted objects, The prediction probability of the model is represented, Representing the adjustable factor.

Description

Two-stage lightweight network-based pleuroperitoneal cavity effusion detection and segmentation method Technical Field The invention relates to the technical field of medical image analysis, in particular to a method for detecting and dividing pleuroperitoneal cavity effusion based on a two-stage lightweight network. Background Hydrothorax and peritoneal fluid are common clinical diseases, which refer to abnormal fluid accumulated in the chest or peritoneal cavity, and are usually related to lesions of organs such as heart, liver, kidney and lung. Timely and accurate detection and segmentation of pleuroperitoneal cavity effusion is of great importance for diagnosing diseases, making treatment schemes and assessing prognosis. Traditional pleuroperitoneal cavity effusion detection relies on manual analysis of images, and doctors manually identify effusion areas according to imaging features. This approach is not only time consuming, but is also limited by physician experience and skill level, which can easily lead to subjectivity and inconsistency in the diagnostic results. With the development of medical imaging technology, particularly the wide application of ultrasonic imaging in the detection of pleuroperitoneal cavity effusion, the imaging method provides more accurate basis for diagnosis. However, the effusion area in the ultrasonic image is often low in contrast, blurred in boundary and represented as a small target in multi-phase or dynamic scanning, so that most of the existing pleuroperitoneal cavity effusion detection technologies still depend on auxiliary tools of manual labeling or semi-automation, and the efficiency is low and the operation is complicated. Therefore, how to improve the automation level of the detection and segmentation of the pleuroperitoneal cavity effusion, lighten the workload of doctors and improve the diagnosis efficiency is a problem to be solved urgently. In recent years, the rapid development of deep learning and computer vision techniques has provided new solutions for medical image analysis. An automatic image analysis method based on a deep learning model such as a Convolutional Neural Network (CNN) has made remarkable progress in the detection and segmentation of pleuroperitoneal cavity effusion. The method can accurately identify the pleuroperitoneal cavity effusion area by automatically learning the image characteristics, so that the detection efficiency and accuracy are improved. However, the existing deep learning model still faces several core problems, namely firstly, difficulty in considering accuracy and calculation efficiency, incapability of meeting the requirements of clinical real-time processing by most high-precision models, secondly, difficulty in clinical deployment caused by excessive algorithm complexity, lack of lightweight design for a low-resource environment, and finally, insufficient robustness of the model, limited generalization capability especially for small targets and weak contrast areas when facing diversified images acquired by different patients and different devices, unstable performance, and influence on universality and reliability in real-world clinical application. Therefore, how to design a lightweight deep learning model suitable for low-power consumption or mobile computing environment, which not only can maintain higher detection and segmentation accuracy, but also can effectively cope with the segmentation challenges of small targets and fuzzy boundaries, is still an important problem faced by the current technology. Disclosure of Invention The invention aims to provide a pleuroperitoneal cavity effusion detection and segmentation method based on a two-stage lightweight network, which aims to solve the problems in the background technology. In order to achieve the above purpose, the invention provides a method for detecting and dividing pleuroperitoneal cavity effusion based on a two-stage lightweight network, which comprises the following steps: S1, acquiring a plurality of groups of pleuroperitoneal cavity ultrasonic images, and marking a detection frame and marking a segmentation mask for a effusion area in the images by a professional to form a training data set; S2, constructing a two-stage deep learning network model, wherein the first stage adopts an SL-YOLO detection model designed for small target detection to position a effusion region, and the second stage adopts a light-weight improved SLNet segmentation model to accurately segment the effusion region; The SLNet segmentation model is a lightweight improved model of a U-Net architecture, the SLNet segmentation model performs feature extraction by adopting a multi-core depth separable convolution module to replace standard convolution in an encoding path, a gating network is introduced between an encoder and a decoder to perform jump connection of adaptive feature screening, and a region perception decoding layer is introduced in a decoding path to fuse global and l