Deep Learning Methods for Cervical Cell Classification with Cytology Image Data
Date
2025-03-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Thesis
Degree Level
Doctoral
Abstract
Advances in biomedical imaging modalities can provide rich information about the properties of normal and cancerous tissues, assisting medical experts in examining and interpreting the current status of cancers. Biomedical image classification plays an integral role in the diagnosis and treatment of various cancers. Traditional methods have some disadvantages, such as time-consuming and labor-intensive, which significantly affect the effective detection of cancer. In response to these issues, it is necessary to apply computer-aided techniques to automatically analyze medical images, so that patients can be diagnosed and treated timely. With the success of deep learning techniques in the field of computer vision, more and more researchers choose machine and deep learning models as major tools for analyzing biomedical images. Leveraging some machine learning-based algorithms, e.g., classification algorithms, is helpful for medical experts to diagnose disease status since computers may discover the patterns that are difficult for humans to see.
Cervical cancer is seriously threatening the lives of women around the world. Fortunately, cervical cancer can be detected and treated as long as it is diagnosed at the early stage. Although traditional methods such as Pap smear test can reduce the death rate of cervical cancer, they are time-consuming and prone to errors. With the emergence of machine and deep learning algorithms, cervical cancer has been diagnosed and treated effectively. However, there are still some challenges in the current detection methods. First, the importance of features of various levels is not fully taken into account in cytology image analysis. Second, the majority of relevant studies only pay attention to the local detailed information while ignoring the global feature information of cells. Moreover, most researchers only focus on the features of images themselves, rather than involving the properties of cervical cells.
To improve these issues, I have developed novel and effective computational methods in Chapters 3-5 of this thesis to automatically analyze cervical cell images, further assisting in the diagnosis and treatment of early cervical cancer.
Chapter 3 develops a new deep convolutional neural network (CNN) that learns feature patterns by applying multiple kernels with different sizes to automatically classify cervical cells from cytology images. Three different basic modules are designed to capture feature information through multiple kernels with different sizes. Afterwards, several such basic modules are stacked to form the cervical cell classification model. Chapter 4 considers the importance of global features. A deep integrated feature fusion module is proposed to synergize local and global cell features more effectively. Based on a parallel architecture with a CNN branch and a transformer branch, the local and global features are extracted simultaneously, enhancing the discriminative ability of cell features. Chapter 5 incorporates supervised contrastive learning with CNN to construct a novel cervical cell classification framework, to guide the deep model to acquire better cell representations. By generating multiple views of cell images with data augmentation to simulate the pattern variations of cervical cells and jointly training the model with our designed hybrid loss, the proposed framework improves the robustness in cervical cell representations and addresses challenges, including the pattern variations in cervical cells caused by the data acquisition process and the misclassification of cells with similar pathological properties.
In summary, this thesis begins with relevant background knowledge that describes the proposed research objectives and pipelines. Subsequently, the thesis presents a comprehensive review of the existing literature on deep learning-based cervical cell image analysis, identifying key findings and challenges in the current research, as well as providing a foundation for the subsequent analysis. Next, this thesis develops several new methodologies for automatic detection of cervical cell abnormalities. A substantial number of comparative experiments demonstrate the effectiveness of the developed methods, which could be viewed as an auxiliary tool to support clinical use for cervical cancer.
Description
Keywords
Cervical cell image classification, Cervical cancer, Convolution neural network, Deep learning, Transformer
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
Biomedical Engineering
Program
Biomedical Engineering