A preoperative identification method for enteric ganglion cells based on CE-Net algorithm

Abstract

The present invention discloses a method for identifying intestinal ganglion cells based on the CE‑Net algorithm, comprising the following steps: S1, collecting image data of pathological sections, and increasing the amount of image data by preprocessing to obtain a training data set; S2, constructing an intestinal ganglion cell recognition model; S3, training the intestinal ganglion cell recognition model in combination with the training data set; S4, calculating a threshold at the inference end of the intestinal ganglion cell recognition model; S5, using the trained intestinal ganglion cell recognition model to identify intestinal ganglion cells in the image to be tested. The present invention can realize rapid and accurate identification of ganglion cells, and reduce the time cost and error rate of manual identification.

Description

Technical Field

[0001] The present invention relates to the technical field of cell identification, and in particular to a preoperative identification method of enteric ganglion cells based on a CE-Net algorithm. Background Art

[0002] Aganglionosis, as the name implies, is a lack of ganglion cells in the intestine. It is a congenital abnormal development of the intestinal nervous system with an incidence of about 1 / 2000-1 / 5000. Due to the lack of intestinal ganglion cells, the diseased intestine is in a spasmodic state and cannot relax, leading to intestinal obstruction. Since the intestinal nervous system develops from the head to the tail in the embryonic period, if the development stops at a certain part of the digestive tract, then the intestinal ganglion cells are lacking beyond that part. Therefore, the lesions of aganglionosis are mostly located at the distal end of the digestive tract, especially in the middle and distal end of the colon, which leads to the clinical characteristics of the disease such as difficulty in defecation, intestinal obstruction, abdominal distension, intestinal infection, etc. Aganglionosis is a functional disease caused by pathological abnormalities. The basic pathological changes include: the absence of ganglion cells in the submucosal and myenteric plexus of the diseased intestine; the proliferation of hypertrophic bile linergic nerve fibers in the submucosal layer and myenteric layer of the intestinal wall without nerve segments, forming the so-called "hypertrophic nerve trunk" structure. Surgical resection of diseased intestine is the main method for clinical treatment of children.

[0003] Intestinal aganglionosis requires surgical treatment, which requires a clear preoperative diagnosis to ensure that there is no misdiagnosis or surgical error. In addition to clinical symptoms and signs, preoperative diagnostic methods mainly include barium enema and rectal biopsy. Barium enema is used to determine the length of the diseased intestinal tract, that is, the severity of the disease. Because intestinal aganglionosis is a pathological disease, preoperative pathological diagnosis is the gold standard. It can be a rectal mucosal biopsy or a rectal (colon) biopsy to determine whether there are no ganglion cells in the submucosal or myenteric plexus. Because the preoperative pathological biopsy is often small, and most cases are transanal rectal mucosal suction biopsies (including the submucosal layer, that is, the submucosal plexus), hematoxylin-eosin staining (H&E) is used to evaluate the characteristics of ganglion cells and plexuses, which requires high film reading skills and experience of pathologists.

[0004] The pathological diagnosis guideline for intestinal aganglionosis recommends: When a mucosal biopsy or full-thickness biopsy is used, it is recommended to use at least one of immunohistochemistry and AchE enzyme histochemistry as an auxiliary diagnostic method. However, in view of the current development status of pediatric surgery and related pathological diagnosis, paraffin sections combined with H&E staining are still the main diagnostic method for pathologists. As long as ganglion cells are found in the pathological sections, intestinal aganglionosis can be ruled out. Intestinal aganglionosis can only be diagnosed if ganglion cells are not found in at least 20 slides. This requires pathologists to spend a lot of working time, have rich accurate film reading technology and the ability to capture signals.

[0005] Using artificial intelligence to quickly identify nerve plexuses and ganglion cells in H&E-stained images of intestinal tissue sections and establish a standard and stable diagnostic method for intestinal aganglionosis can reduce dependence on experienced pathologists and reduce the workload of pathologists. It can also help to promote it to grassroots units, achieve homogeneity of pathological diagnosis between hospitals of different levels, and improve the overall quality of medical care.

[0006] In recent years, deep learning methods based on U-Net have been continuously proposed in the field of medical imaging. However, a common problem of U-Net and its variants is that continuous pooling and strided convolution reduce feature resolution, resulting in the loss of some spatial information. Although this invariance is beneficial for classification or object detection tasks, it is often not conducive to dense prediction tasks that require detailed spatial information. Intuitively, maintaining high-resolution feature maps in the middle layer can improve segmentation accuracy. However, this increases the size of the feature map, which is not conducive to accelerating network training and reducing the difficulty of network optimization.

[0007] Based on the above-mentioned defects of the U-Net method in the field of medical images, it is necessary to find a compromise between accelerating training and maintaining high resolution. Generally, U-Net can be regarded as an encoder-decoder structure, where the encoder aims to gradually reduce the spatial dimension of the feature map and capture more high-level semantic features; the decoder aims to restore the target detail information and spatial dimension. Therefore, capturing higher-level features in the encoder and retaining more spatial information in the decoder to improve the performance of image segmentation for application in the recognition of enteric ganglion cells and obtaining higher recognition effects are the problems that this patent aims to solve. Summary of the invention

[0008] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a preoperative identification method for enteric ganglion cells based on the CE-Net algorithm, which can accurately and quickly identify enteric ganglion cells, thereby assisting pathologists in preoperative diagnosis, improving the speed and accuracy of identification, and reducing the burden on pathologists.

[0009] Technical solution: In order to achieve the above-mentioned invention object, the present invention provides a method for preoperative identification of enteric ganglion cells based on the CE-Net algorithm, comprising the following steps:

[0010] S1, collecting image data of pathological sections and increasing the amount of image data through preprocessing to obtain a training data set;

[0011] S2, construct an enteric ganglion cell recognition model;

[0012] S3, train the enteric ganglion cell recognition model based on the training dataset;

[0013] S4, threshold calculation at the inference end of the enteric ganglion cell recognition model;

[0014] S5, the intestinal ganglion cells are recognized in the test image through the trained intestinal ganglion cell recognition model.

[0015] Furthermore, in the present invention: S1 further comprises the following steps:

[0016] S1-1, perform data enhancement on image data;

[0017] S1-2, randomly select 100 images without ganglion cells and generate all-black labels to expand the dataset and make full use of the provided images;

[0018] S1-3, increasing the number of images through pseudo-labeling strategy.

[0019] Furthermore, in the present invention: the data enhancement includes: adopting a resizing strategy for the image, adjusting the brightness, contrast, saturation, and color enhancement of the image in order to improve the generalization ability of the model;

[0020] Furthermore, in the present invention: the enteric ganglion cell recognition model is constructed and optimized based on the CE-Net network of the U-Net architecture. The CE-Net network serves as the basic model of the enteric ganglion cell recognition model, including a feature encoding module, a context extraction module and a feature decoding module.

[0021] Furthermore, in the present invention: the optimization also includes: optimizing the backbone network, deepening the network layer and using Efficientnet as the encoder in the model; embedding the modules of the channel attention mechanism SENet and the self-attention mechanism DANet into the structure; and adopting the feature pyramid pooling of PSPNet.

[0022] Furthermore, in the present invention: S3 further includes the following steps:

[0023] S3-1, using the adam optimizer;

[0024] S3-2, adaptive learning rate adjustment strategy, when epoch_loss is greater than best_loss three times, the learning rate is divided by 2;

[0025] S3-3, set the loss function. The loss function adopts the combination of BCEloss+DICEloss+FOCALloss, where BCEloss is a binary cross entropy loss function, DICEloss is used to improve the DICE score of the enteric ganglion cell recognition model, and FOCALloss is used to balance positive and negative samples.

[0026] Furthermore, in the present invention: the calculation of the threshold value also includes: randomly sampling training set pictures, calculating the best threshold value and taking the average value as the threshold value.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: a recognition model for ganglion cells in the intestine is constructed based on the CE-Net algorithm, which can realize rapid and accurate recognition of ganglion cells and reduce the time cost and error rate of manual recognition.

[0028] (1) The present invention uses the CE-Net algorithm to obtain more advanced information and retain spatial information for preoperative pathological auxiliary diagnosis of intestinal aganglionosis; the model has fast operation speed, few parameters, few dependencies, and is easy to deploy;

[0029] (2) The network structure proposed in this paper adds a context extractor to the traditional encoder-decoder structure to reduce the information loss caused by pooling and convolution; the context extractor uses DAC blocks (cavity convolution) and RMP blocks (residual multi-core pooling), which are combined with the encoder-decoder structure to capture more high-level features and retain more spatial information, making medical image segmentation more accurate;

[0030] (3) DACblock (Dense Atrous Convolution) uses different rate holes to increase the receptive field, as well as the convolution strategy of the Inception module, while RMP (Residual Multi-kernel pooling) increases the receptive field and regional information by expanding the pooling area and then upsampling for splicing. Specifically, the original U-Net architecture captures multi-scale features by adopting continuous 3*3 convolution and pooling operations during the encoding process, while the DAC module proposed by CE-Net can capture broader and deeper semantic features by injecting four cascaded branches with multi-scale hole convolutions. In this module, shortcut connections are used to prevent the gradient vanishing problem. In addition, inspired by spatial pyramid pooling, CE-Net also proposes a residual multi-kernel pooling (RMP) module. This RMP module further encodes the multi-scale contextual features of the target extracted from DAC by adopting pooling operations of various sizes without additional learning weights. In summary, the DAC module uses multi-scale hole convolutions to extract rich feature expressions, and then the RMP module uses multi-scale pooling operations to further obtain contextual information. Combining the DAC module and the RMP module with the encoder-decoder structure, the context encoding network relies on the DAC module and the RMP module to capture more abstract features and retain more spatial information to improve the performance of medical image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the overall process of the method proposed in the present invention;

[0032] Figure 2 This is a schematic diagram of normal ganglion cells in the myenteric plexus;

[0033] Figure 3 Schematic diagram of normal ganglion cells in the submucosal plexus. DETAILED DESCRIPTION

[0034] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0035] The present invention may be implemented in many different forms and should not be considered limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0036] like Figure 1 As shown, the present invention proposes a preoperative identification method for enteric ganglion cells based on the CE-Net algorithm, and the method specifically comprises the following steps:

[0037] S1, collecting image data of pathological sections and increasing the amount of image data through preprocessing to obtain a training data set;

[0038] The preprocessing of image data also includes the following steps:

[0039] S1-1, data enhancement is performed on the image data; specifically, for the segmentation under the analysis of the pathological slice, which does not belong to ultra-fine segmentation, a resizing strategy is adopted for the image in this embodiment. To improve the generalization ability of the model, data enhancement also includes adjusting the brightness, contrast, saturation, and color enhancement of the image;

[0040] S1-2, in order to expand the dataset and make full use of the provided images, 100 images without ganglion cells were randomly selected and all-black labels were generated to expand the dataset and make full use of the provided images.

[0041] S1-3, increase the number of images through pseudo-labeling strategy; take the ganglion region from the image with ganglion cells and graft it onto the provided image without ganglion cells.

[0042] S2, construct an enteric ganglion cell recognition model;

[0043] Specifically, the enteric ganglion cell recognition model in this embodiment is constructed based on the CE-Net network of the U-Net architecture. The CE-Net network, as the basic architecture of the enteric ganglion cell recognition model, includes three modules: a feature encoding module, a context extraction module, and a feature decoding module.

[0044] Furthermore, the enteric ganglion cell recognition model used in this embodiment improves the structure of the CE-Net network, and the improvements include the following aspects:

[0045] The backbone network is optimized by deepening the network layer and using the lighter and more efficient Efficientnet as the encoder in the model;

[0046] Embed the modules of channel attention mechanism SENet and self-attention mechanism DANet into the structure;

[0047] Feature pyramid pooling of PSPNet is used.

[0048] S3, train the enteric ganglion cell recognition model based on the training dataset;

[0049] S3-1, using the Adam optimizer.

[0050] S3-2, adaptive learning rate adjustment strategy, when epoch_loss is greater than best_loss three times, the learning rate is divided by 2.

[0051] S3-3, set the loss function. The loss function adopts the combination of BCEloss+DICEloss+FOCALloss. BCEloss is a binary cross entropy loss function, DICEloss is to improve the DICE score of the model, and FOCALloss is to balance the positive and negative samples.

[0052] S4, The inference side of the enteric ganglion cell recognition model calculates the segmentation threshold.

[0053] Specifically, a threshold is required for inference, and the result is classified into pixels by the threshold. That is, a probability is predicted for each pixel in the image, and then the probability is mapped to 0 or 1 according to the threshold.

[0054] This embodiment randomly samples 30 training set images, calculates the best threshold and then takes the average value as the segmentation threshold used in the present invention. In order to utilize the non-fixed orientation of pathological images, the strategy of flipping the image test is adopted during the test.

[0055] Among them, the flipped image is used for testing to take advantage of the fact that the pathological image has a non-fixed direction and to test the robustness of the model to direction flipping. After obtaining the segmentation threshold, the original image is first used for testing, and then the image is flipped for testing.

[0056] The minimum value of the threshold range is 0, and the maximum value is 1. Generally, 0.5 is predicted to be the threshold. The segmentation threshold selected in this embodiment is 0.4918.

[0057] S5, the intestinal ganglion cells are recognized in the test image through the trained intestinal ganglion cell recognition model.

[0058] Among them, the image to be tested is a stained intestinal pathological section image.

[0059] It should be noted that the above-described embodiments only express some implementation methods of the present invention, and the description thereof cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the concept of the present invention, and these should all fall within the scope of protection of the present invention.

Claims

1. A preoperative identification method for enteric ganglion cells based on CE-Net algorithm, characterized in that: The following steps are involved: S1, collecting image data of pathological sections and increasing the amount of image data through preprocessing to obtain a training data set; S2, construct an enteric ganglion cell recognition model; S3, train the enteric ganglion cell recognition model based on the training dataset; S4, threshold calculation at the inference end of the enteric ganglion cell recognition model; S5, using the trained enteric ganglion cell recognition model to identify enteric ganglion cells in the test image; The S1 further comprises the following steps: S1-1, perform data enhancement on image data; S1-2, randomly select 100 images without ganglion cells and generate all-black labels to expand the dataset and make full use of the provided images; S1-3, increasing the number of images through pseudo-labeling strategy; The data enhancement includes: adopting a resizing strategy for the image, adjusting the brightness, contrast, saturation, and color enhancement of the image to improve the generalization ability of the model; The enteric ganglion cell recognition model is constructed and optimized based on the CE-Net network of the U-Net architecture. The CE-Net network is used as the basic model of the enteric ganglion cell recognition model, including a feature encoding module, a context extraction module and a feature decoding module. The optimization also includes: optimizing the backbone network, deepening the network layer and using Efficientnet as the encoder in the model; embedding modules of the channel attention mechanism SENet and the self-attention mechanism DANet into the structure; and using feature pyramid pooling of PSPNet.

2. The method for preoperative identification of enteric ganglion cells based on the CE-Net algorithm according to claim 1, characterized in that: The S3 further comprises the following steps: S3-1, using the adam optimizer; S3-2, adaptive learning rate adjustment strategy, when epoch_loss is greater than best_loss three times, the learning rate is divided by 2; S3-3, set the loss function. The loss function adopts a combination of BCEloss, DICEloss and FOCALloss. Among them, BCEloss is a binary cross entropy loss function, DICEloss is used to improve the DICE score of the enteric ganglion cell recognition model, and FOCALloss is used to balance positive and negative samples.

3. The method for preoperative identification of enteric ganglion cells based on the CE-Net algorithm according to claim 2, characterized in that: The calculating threshold value also includes: randomly sampling training set pictures, calculating the best threshold value and taking the average value as the threshold value.

Images