Diagnosis of cervical precancerous lesions based on multimodal feature changes

Highlights

• A diagnosis method of cervical precancer based on the preacetic and postacetic acid test cervical images was developed.

• A deep learning network was used to extract features and classify the registered cervical images.

• The proposed method explored different strategies for classification.

Abstract

To realize the automatic diagnosis of cervical intraepithelial neoplasia (CIN) cases by preacetic acid test and postacetic acid test colposcopy images, this paper proposes a method of cervical precancerous lesion diagnosis based on multimodal feature changes. First, the preacetic acid test and postacetic acid test colposcopy images were registered based on cross-correlation and projection transformation, and then the cervical region was extracted by the k-means clustering algorithm. Finally, a deep learning network was used to extract features and classify the preacetic acid test and postacetic acid test cervical images after registration. Finally, the proposed method achieves a classification accuracy of 86.3%, a sensitivity of 84.1%, and a specificity of 89.8% in 60 test cases. Experimental results show that this method can make better use of the multimodal features of colposcopy images and has lower requirements for medical staff in the process of data acquisition. It has certain clinical significance in cervical cancer precancerous lesion screening systems.

Introduction

Cervical cancer is the fourth leading cause of death in women, with a high incidence rate and mortality [1,2]. It is estimated that there will be 570000 cases and 311000 deaths worldwide in 2018 [1]. At the same time, it is the second most common cancer for women worldwide [3], with a cure rate of nearly 100% in the early stages and 20%–50% in the late stages [[4], [5], [6], [7], [8]]. Therefore, early screening of cervical cancer is of great significance to the prevention and treatment of cervical cancer.

The precancerous lesions of cervical cancer are mainly cervical intraepithelial neoplasia (CIN), which can be divided into low-grade squamous intraepithelial lesions (LSIL, including normal and CIN1) and high-grade squamous intraepithelial lesions (HSIL, including CIN2 and CIN3) [[9], [10], [11]]. Among them, CIN1, CIN2, and CIN3 are the three CIN grades stipulated by the World Health Organization. Based on the existing medical environment and conditions, LSIL only needs conservative observation and treatment. By improving the living habits and environmental hygiene of patients, LSIL can often be eliminated by the immune system [6,12], while HSIL needs timely diagnosis and treatment by doctors.

At present, cervical cancer screening methods mainly include human papillomavirus detection [13,14], cervical smear [13,15] and acetic acid testing under colposcopy, but the most commonly used cervical screening method is mainly the acetic acid test under colposcopy [[16], [17], [18], [19]]. By smearing 3%–5% acetic acid on the surface of the cervix, digital colposcopy equipment is used to take pictures to record the changes in the cervical region, and then experienced medical staff judged whether there was a vinegar white reaction in the cervical area to analyze the possibility of CIN. However, the acetic acid test, as a visual diagnosis method, has strong subjectivity in the whole diagnosis process, which requires medical staff to have rich clinical experience, which often lacks medical resources in areas. At the same time, cervical cancer is a large-scale screening disease, and its workload is huge, which, to a certain extent, improves the possibility of missed diagnosis and misdiagnosis of medical staff.

To solve the above problems, an increasing number of researchers have begun to use computer technology to analyze colposcopy images. Through computer technology, the image features are extracted manually or automatically, and finally, different classifiers are used to judge the CIN level. Xu et at [12]. Manually extracted the following three complementary histogram features from colposcopy images: the L*a*b color histogram, gradient histogram, and local binary histogram. Then, SVM was used to classify LSIL and HSIL. Finally, the average accuracy of classification was 77.2%. Asiedu et at [20]. extracted the texture and color features of colposcopy images from acetic acid experiments and iodine experiments, respectively, used an SVM classifier to classify CIN-positive and CIN-negative samples, and finally obtained a classification accuracy of 80%.

With the rise of artificial intelligence (AI), automatic diagnosis of lesions has gradually become a trend. At present, convolutional neural networks have been widely used in image diagnosis and have achieved good results [[21], [22], [23], [24], [25], [26]]. For example, the classification of breast cancer [[27], [28], [29], [30]], bladder cancer [31], lung cancer [32,33] and so on. Therefore, the application of AI technology in colposcopy image analysis is an effective way to realize the automatic diagnosis of cervical cancer and improve the accuracy of CIN classification. Hu et at [34]. proposed a visual assessment method based on deep learning to automatically identify cervical precancerous cancer or cancer. Fast R–CNN is used to train the colposcopy image for the cervical locator, and then CNN is used to train the cervical region for cervical precancerous or cancer classification. The algorithm finally achieves the experimental results with an AUC of 0.91. Yue et at [35]. proposed a method to automatically predict the CIN level of sequential colposcopy images using LSTM based on multistate CNN features. The algorithm uses CNN to extract the features of colposcopy images at different times and states and then uses the LSTM network to further process the extracted features. Finally, the obtained feature results are cascaded, and the final classification processing is carried out. Finally, a classification accuracy of 96.13% was achieved. Li et at [36]. proposed a computer-aided system based on sequential colposcopy images. In this algorithm, ResNet50 is used to extract the cervical region from colposcopy images at different times, and then separate feature codes are used to extract the cervical region features. Finally, a graph convolution network is constructed to fuse the extracted features and classify the positive and negative features. Finally, a classification accuracy of 78.33% was achieved.

According to the existing research data, most of the research on cervical precancerous lesion recognition is based on single colposcopy images after acetic acid experiments, and this single piece of information can lead to difficulty in reflecting changes in cervical vinegar white preacetic acid tests and postacetic acid tests [12,20]. Therefore, this paper proposes a diagnostic method for cervical precancerous lesions based on multimodal feature changes, which can solve these problems. In the first step, the preacetic acid test and postacetic acid test colposcopy images were used for a series of image preprocessing, such as image registration and cervical region extraction. In the second step, transfer learning is used to extract multimodal features from cervical images. The third step is to train the extracted image features to obtain the final classification results.