Search

Because cervical histopathology image analysis plays a very importation role in the cancer diagnosis and medical treatment processes, since the 1980s, more and more effective machine vision techniques are introduced and applied in this field to assist histopathologists to obtain a more rapid, stable, objective, and quantified analysis result. To discover the inner relation between the visible images and the actual diseases, a variety of machine vision techniques are used to help the histopathologists to get to know more properties and characteristics of cervical tissues, referring to artificial intelligence, pattern recognition, and machine learning algorithms. Furthermore, because the machine vision approaches are usually semi- or full-automatic computer based methods, they are very efficient and labour cost saving, supporting a technical feasibility for cervical histopathology study in our current big data age. Hence, in this article, we comprehensively review the development history of this research field with two crossed pipelines, referring to all related works since 1988 till 2020. In the first pipeline, all related works are grouped by their corresponding application goals, including image segmentation, feature extraction, and classification. By this pipeline, it is easy for histopathologists to have an insight into each special application domain and find their interested applied machine vision techniques. In the second pipeline, the related works on each application goals are reviewed by their detailed technique categories. Using this pipeline, machine vision scientists can see the dynamic of technological development clearly and keep up with the future development trend in this interdisciplinary field.

In this paper, a novel multilayer hidden conditional random fields (MHCRFs)-based cervical histopathology image classification (CHIC) model is proposed to classify well, moderate and poorly differentiation stages of cervical cancer using a weakly supervised learning strategy. First, the color, texture, and deep learning features are extracted to represent the histopathological image patches. Then, based on the extracted features, artificial neural network, support vector machine, and random forest classifiers are designed to calculate the patch-level classification probabilities. Third, effective classifiers are selected to generate unary and binary potentials. At last, using the generated potentials, the final image-level classification results are predicted by our MHCRF model, and an overall accuracy around 77.32% is obtained on six practical cervical histopathological image datasets with more than 600 immunohistochemical (IHC) stained samples. Among the six test accuracies, the highest reaches 88%. Furthermore, we also test our MHCRF method with a gastric hematoxylin-eosin (HE) stained histopathological image dataset including 200 images for an extended experiment, and achieve an accuracy of 93%.

To apply topological information to solve a cervical histopathology image clustering (CHIC) problem, a graph based unsupervised learning (GBUL) approach is proposed in this paper. First, the GBUL method applies color features and k-means clustering to carry out a first-stage “coarse” clustering. Then, a skeletonization based node generation (SBNG) approach is introduced to approximate the distribution of cervical cell nuclei. Thirdly, based on the SBNG nodes, multiple graphs are constructed. Next, graph features are extracted based on the constructed graphs. Finally, k-means clustering is used again for the second-stage clustering. In the experiment, a practical Hematoxylin–eosin staining cervical histopathology image dataset with 40 whole-slide imaging images is tested, obtaining a promising CHIC result and showing a huge potential in the cancer risk prediction field.

In order to apply the important topological information to solve a Cervical Histopathology Image Clustering (CHIC) problem, a Graph Based Unsupervised Learning (GBUL) approach is proposed in this paper. First, the GBUL method applies color features and k-means clustering for a first-stage “coarse” clustering. Then, a Skeletonization Based Node Generation (SBNG) approach is introduced to approximate the distribution of cervical cell nuclei. Thirdly, based on the SBNG nodes, a minimum spanning tree graph is constructed. Next, graph features and additional geometrical features are extracted based on the constructed graph. Finally, the k-means clustering is applied again for the second-stage clustering. In the experiment, a practical cervical histopathology image dataset with ten whole scanned images is tested, obtaining a promising CHIC result and showing a huge potential in the cancer risk prediction field.

Because Breast Histopathology Image Analysis (BHIA) plays a very important role in breast cancer diagnosis and medical treatment processes, more and more effective Machine Learning (ML) techniques are developed and applied in this field to assist histopathologists to obtain a more rapid, stable, objective, and quantified analysis result. Among all the applied ML algorithms in the BHIA field, Artificial Neural Networks (ANNs) show a very positive and healthy development trend in recent years. Hence, in order to clarify the development history and find the future potential of ANNs in the BHIA field, we survey more than 60 related works in this paper, referring to classical ANNs, deep ANNs and methodology analysis.

In recent years, Transfer Learning makes a great breakthrough in the field of machine learning, and the use of transfer learning technology in Cervical Histopathology Image Classification (CHIC) becomes a new research domain. In this paper, we propose an Ensembled Transfer Learning (ETL) framework to classify well, moderately and poorly differentiated cervical histopathology images. In this ETL framework, Inception-V3 and VGG-16 based transfer learning structures are first built up. Then, a fine-tuning approach is applied to extract effective deep learning features from these two structures. Finally, a late fusion based ensemble learning strategy is designed for the final classification. In the experiment, a practical dataset with 100 VEGF stained cervical histopathology images is applied to test the proposed ETL method in the CHIC task, and an average accuracy of 80% is achieved.