Early screening of cervical cancer based on tissue Raman spectroscopy combined with deep learning algorithms

Cervical cancer is the most common reproductive malignancy in the female reproductive system. Studies have shown that it has the 2nd highest incidence and 4th highest mortality rate among female tumors worldwide, with approximately 90% of the mortality occurring in developing countries [1,2]. Cervical carcinomatosis is a relatively slow process, with patients taking roughly 5 to 8 years to progress from HPV infection to cause CIN lesions, then from LSIL to HSIL via cervical intraepithelial neoplasia, and finally to cervical cancer [3]. Cervical cancer can be cured if detected and treated appropriately in time, thus effectively reducing morbidity and mortality, so early screening for cervical cancer is especially critical and significant. However, all currently commonly used clinical screening methods have shortcomings, for example, Pap smear tests have high false negative results [4,5]; cytology-based methods have high specificity but low sensitivity [6]; and HPV DNA tests are highly sensitive but low specificity and expensive [7]. Therefore, there is an urgent need to develop a stable, efficient and rapid diagnostic method to assist in the diagnosis and pathological study of cervical cancer.

Raman spectroscopy is an optical detection technique based on the inelastic scattering of light by vibrating molecules, which has the advantages of rapid, non-invasive, high sensitivity and high specificity. In the past decades, Raman spectroscopy has been widely used for early screening and diagnosis of various diseases, such as prostate cancer, ovarian cancer, pancreatic cancer, brain tumors, oral cancer and cervical cancer [8], [9], [10], [11]. In the current study, spectroscopic techniques have been combined with multivariate statistics to achieve results in the field of medical diagnosis of diseases related to cervical disorders. For example, in 2005, Faolain et al. [12] directly compared Raman spectroscopy and synchrotron infrared (SR-IR) spectroscopy on parallel cervical cancer samples. In 2014, Rashid et al. [13] evaluated the potential of Raman microspectroscopy for elucidation of the biochemical changes associated with the pre-malignant stages of cervical cancer. In 2015, Lyng, et al. [14] discussed the recent advances and challenges for cervical-cancer screening and diagnosis and offers some perspectives for the future. In 2016, Ramos et al. [15] used principal component analysis (PCA) to distinguish normal and abnormal ThinPrep according to the biochemical fingerprints of cells Samples. Principal component analysis-linear discriminant analysis (PCA-LDA) was used to establish classification model based on CIN or SIL terms. In 2020, Karunakaran et al. [16] established for the first time a diagnostic system based on label-free surface-enhanced Raman spectroscopy using exfoliated cell samples as a way to differentiate between healthy individuals, precancerous lesions and squamous cell cervical cancer in the invasive zone. They developed PCA-LDA and PCA-SVM classification models and determined the average diagnostic accuracy of the three to be 94%, 74% and 92%, respectively. In the same year, Chen et al. [17] proposed an airPLS-PLS-GA-SVM efficient HPV screening method by combining Raman spectroscopy of cervical secretions and pattern recognition algorithm, in which the total discriminatory accuracy of the experiment was 98.6%.

There have been many significant research results on various frameworks incorporating attention mechanisms, and novel models based on neural networks are emerging. The attention mechanism can reduce the computational burden of processing high-dimensional input data on the one hand, and reduce the data dimensionality by structurally selecting a subset of the input. On the other hand, artifacts can be removed to allow the task processing system to focus more on finding information in the input data that is significantly relevant to the current output, thus improving the quality of the output [18]. Attentional mechanisms were first proposed in the field of visual images [19]. subsequently, Bahdanau et al. applied them to the field of NLP [20].

In this study, Raman spectroscopy combined with deep learning algorithm was used to classify and identify tissue samples from cervicitis, Low-grade Squamous Intraepithelial Lesion (LSIL), High-grade Squamous Intraepithelial Lesion (HSIL), Well Differentiated Squamous Cell Carcinoma (WDSCC), Moderately Differentiated Squamous Cell Carcinoma(MDSCC), Poorly Differentiated Squamous Cell Carcinoma(PDSCC) and cervical adenocarcinoma. The attention mechanism efficient channel attention network (ECANet) module and squeeze-and-excitation network (SENet) module were combined with convolutional neural network (CNN) and residual neural network (ResNet), respectively, to construct a diagnostic model for multi-classification of cervical cancer. The experimental results show that the attention mechanism module can lead to better diagnostic performance of the model.

留言 (0)

沒有登入
gif