CN-114998264-B - Automatic detection method for TCT smear

CN114998264BCN 114998264 BCN114998264 BCN 114998264BCN-114998264-B

Abstract

The invention discloses an automatic detection method of a TCT smear, which comprises the steps of constructing DeepTCT, carrying out iterative training on DeepTCT, namely searching cervical cells detected by DeepTCT in each image of a test set, if the classification score corresponding to the cervical cells is larger than a set threshold value, reserving labels of the cervical cells, otherwise, discarding the labels, putting a sample reaching the set threshold value into a training set, carrying out DeepTCT training together with the existing image of the training set, applying the trained DeepTCT to the test set, outputting classification and positioning results of the cervical cells, repeating the iterative training for set times, and applying the final DeepTCT to morphological detection of the cervical cells. According to the invention, through repeated iterative training, the detection performance of DeepTCT on morphological characteristics of cervical cells is remarkably improved, and compared with the existing method, the method provided by the invention has higher mAP and mAR.

Inventors

CHEN TINGMEI
LIU RAN
YAO MENGLI
CHEN XIN
LI FANG
TONG XUAN
WANG ZHENG
YI LIN

Assignees

重庆医科大学
重庆大学

Dates

Publication Date: 20260505
Application Date: 20220602

Claims (2)

1. The automatic TCT smear detection method is characterized by comprising the following steps: 1) Constructing a cell detection model DeepTCT, wherein the model DeepTCT comprises a main network PVTv with a characteristic pyramid network for extracting characteristic images, a region suggestion network for generating preliminary region suggestions, three detection branches for obtaining classification and positioning of cervical cells and an online difficult sample mining module; The three detection branches have the same structure, and each detection branch comprises RoIAlign layers, an average pooling layer connected with RoIAlign layers, a flat layer connected with the average pooling layer, a first full-connection layer connected with the flat layer, a second full-connection layer connected with the first full-connection layer and used for frame regression, and a third full-connection layer connected with the first full-connection layer and used for frame classification; The output of PVTv is taken as the input of the area suggestion network, the output of PVTv and the output of the area suggestion network together form an area of interest of any size and are taken as the input of a first detection branch, the output of a second full-connection layer of the first detection branch is taken as the input of a second detection branch, and the output of a second full-connection layer of the second detection branch is taken as the input of a third detection branch; The on-line difficult sample mining module comprises an average pooling layer, a flutten layer connected with the average pooling layer, a first full-connection layer connected with the flutten layer, a second full-connection layer connected with the first full-connection layer and used for frame regression, and a third full-connection layer connected with the first full-connection layer and used for frame classification, wherein the average pooling layer, the flutten layer, the first full-connection layer, the second full-connection layer and the third full-connection layer in the on-line difficult sample mining module and the third detection branch have the same structure, and the average pooling layer, the flutten layer, the first full-connection layer, the second full-connection layer and the third full-connection layer in the on-line difficult sample mining module share weights; 2) DeepTCT is trained and tested, comprising the steps of: a) Basic training and testing, namely performing data enhancement and class equalization on TCT images in a training set, inputting DeepTCT images in the training set for training, finally applying a trained network DeepTCT to a testing set, and outputting classification and positioning results of cervical cells; b) Iterative training and testing of the basic trained DeepTCT, which in turn includes the steps of: b1 Searching cervical cells detected by DeepTCT in each image of the test set, if the classification score corresponding to the cervical cells is greater than a set threshold, retaining the label of the cervical cells, otherwise discarding the label; b2 Placing the sample which is generated by DeepTCT and has the classification score reaching the set threshold value into a training set, and using the sample with the existing images of the training set for the training of DeepTCT; b3 Applying the trained DeepTCT to the test set, and outputting classification and positioning results of cervical cells; b4 Repeating steps b 1) -b 3) for a preset iteration number; 3) Applying DeepTCT obtained in the step 2) to morphological detection of cervical cells.
2. The method for automated TCT smear detection according to claim 1, wherein the threshold value set in the step b 1) is 0.8.

Description

Automatic detection method for TCT smear Technical Field The invention belongs to the technical field of neural networks, and particularly relates to an automatic detection method for TCT smear. Background The cervical cell growth process and morphological changes are closely related to the occurrence and development of cervical cancer. Cervical cytology is an effective cervical cancer screening tool. Cervical cytology screening is typically accomplished by a pathologist under a microscope, such as conventional pap smear and liquid-based thin layer cytology (Thinprep Cytologic Test, TCT). Pap smear is usually classified into five categories, while TCT is classified into 7 categories by using TBS reporting. Typically, pathologists determine whether cells are abnormal based on morphological features of cervical exfoliated cells. These morphological features include nucleus size, morphology, shade of staining, nuclear to plasma ratio, etc., which play a very important role in the decision of the physician. The samples obtained by the traditional Papanicolaou smear method are often affected by blood, mucus and inflammation, so that the samples are fuzzy, and the judgment of a doctor on morphological characteristics is affected, so that the screening result is inaccurate. TCT is not limited by these factors, and the quality of the sample and the detection rate of abnormal cervical cells are greatly improved. At the same time, TCT may also find some precancerous lesions and microbial infections, such as mold, trichomonas, viruses and chlamydia, etc. It can be said that the advent of TCT has led to a new height of cervical cytology screening. However, TCT screening is currently mainly manual, which is time-consuming and laborious, and accuracy is also susceptible to the level of skill of the doctor and subjective emotion. Therefore, it is necessary to implement automatic screening of cervical cells by means of a machine, thereby assisting medical diagnosis. In order to realize automatic screening of cervical cytology, the machine needs to automatically acquire the characteristics through training, and then the classification and positioning of cervical cells are completed according to the characteristics. Conventional cell morphology feature detection methods include threshold-based methods, cluster-based methods, contour-based methods, conventional machine learning methods, and the like. Due to the complexity of the morphological features of cervical cells, the performance of these methods often fails to meet the clinical practical requirements. In recent years, attempts have been made to apply deep learning to cervical cytology screening and to achieve superior results over conventional algorithms in most tasks. Among them, the Region-based deep convolutional neural network (Region-based Deep Convolutional Neural Networks, R-DCNN) has become the mainstream network model in cervical cytology screening because of its superior performance. The network obtains candidate areas firstly, and then classifies the candidate areas and carries out frame regression. The method has the advantages that the method not only can realize the content identification and classification of the images, but also can more accurately solve the positioning problem of the object to be detected. However, when it is applied to cervical cytology screening, performance is still difficult to meet the requirements of wide clinical application. The Pyramid Vision Transformer (PVT) technology, which has emerged in recent years, offers the possibility of performance improvement for target detection. How to construct an automated screening model with more excellent performance using this technique (including PVTv a 2) remains a challenge. Disclosure of Invention In view of the above, the present invention is to provide an automatic detection method for TCT smear to solve the technical problems of classifying and positioning cervical cells. The automatic detection method of the TCT smear comprises the following steps: 1) Constructing a cell detection model DeepTCT, wherein the model DeepTCT comprises a backbone network PVTv with a characteristic pyramid network for extracting characteristic images, a region suggestion network for generating preliminary region suggestions, three detection branches for obtaining classification and positioning of cervical cells and an online difficult sample mining module, The three detection branches have the same structure, and each detection branch comprises RoIAlign layers, an average pooling layer connected with RoIAlign layers, a flat layer connected with the average pooling layer, a first full-connection layer connected with the flat layer, a second full-connection layer connected with the first full-connection layer and used for frame regression, and a third full-connection layer connected with the first full-connection layer and used for frame classification; The output of PVTv is taken as the input of the area