Global Research Trends of Artificial Intelligence on Histopathological Images: A 20-Year Bibliometric Analysis

Cancer is a widespread disease that causes major death in the world. In 2020, more than 19 million cases of cancer were confirmed worldwide, including over 10 million deaths from cancer [1]. The early diagnosis and treatment for cancer are essential to reduce cancer deaths in the world. Histopathological diagnosis is the gold standard for tumor diagnosis, grade, and classification. Although CT, MRI, and other imaging methods for tumor examination are provided clinically, cancer can only be diagnosed by biopsy and histopathological examination. Therefore, accurate histopathological diagnosis is critical for cancer screening and diagnosis. Histopathological examination is a highly subjective task that relies on professional knowledge and experience to evaluate tumor tissue structure and cell morphology in pathological sections. Unfortunately, pathologists with highly professional experience are often in short supply. Objective and accurate artificial intelligence (AI) histopathological methods could reduce the clinical misdiagnosis of histopathology caused by subjectivity and the lack of professionals [2].

AI is a technology that utilizes computer technology and algorithms to imitate human intelligence. With the rapid development of computer technology, AI-based machine learning and deep learning methods have been widely used in image processing [3]. As a data-driven approach, AI learns features from datasets relevant to downstream tasks. Digital pathology technology enables large-scale digitization of histopathological slices, and the massive growth of data of histopathological images (HI) accelerates the rapid development of AI in this field. AI has been used in early cancer diagnosis [4,5], cancer mutation detection [6,7], nuclear segmentation [8,9], prognosis evaluation [10,11], etc. The combination of AI and pathologists is a conducive way to solve the lack of professional pathologists in some remote regions [12,13,14]. A large number of papers have been published on AI applications in the fields of HI, so it is worthwhile to quantitatively analyze the papers in this field to grasp the research status and research trends as a whole.

Bibliometrics is a method used to quantitatively analyze numerous documents in a certain field. By counting the authors, countries, journals, and citation relationships of published literature in the field, the bibliometrics can visualize the connection and structure of research topics among the literature to determine current research hotspots and future development trends [15]. There are many studies reviewed on the topic of AI in HI [16,17,18,19], but only one bibliometric study about breast cancer in histopathological image classification [20]. The study is limited to the classification task and breast cancer types. It is necessary to conduct a bibliometric analysis of multi-cancer and multi-tasks on the application of AI in HI. The research themes and emerging trends can be objectively displayed in the field through quantitative analysis of the literature data.

This study is the first bibliometric study to provide a comprehensive view of the application of AI in HI. We collected the publications from 2001 to 2021 in the Web of Science collection database, then counted and visualized multiple indicators by VOSviewer and CiteSpace. This study summarized the prevailing topics of cancer types, methods, and development trends of AI in HI, and provides a reference for future researchers to grasp key points.

In this study, we conducted a bibliometric analysis of 2844 documents to understand the research status, research hotspots, and research development trends in the application of AI in HI analysis. In the past 20 years, the number of publications has increased year after year, especially sharply in the past five years. Breast cancer, prostate cancer, colorectal cancer, and lung cancer are the four most studied cancer types. The USA is the most influential and productive country with far more citations than any other country, and it has collaborated extensively with multiple countries. American researchers and institutions have carried out artificial intelligence pathology research for the longest and have published many high-quality articles. Meanwhile, countries and institutions should strengthen cooperation with each other to promote advanced research and interdisciplinary frontier research.

With comprehensive keyword co-occurrence analysis and reference co-citation analysis, we summarize the challenges and main directions of AI in HI analysis. On the one hand, classification based on deep learning and nuclei segmentation based on deep learning may be the two main research directions at present. Usually, deep learning methods require large datasets with annotation to reduce overfitting and improve generalization capability. On the other hand, articles related to transfer learning and self-supervised learning were published and updated rapidly, indicating these two directions could be potential hotspots in the future.

Classification based on deep learning. Classification is aimed at diagnosing cancer and identifying cancerous tissue. AI methods will develop the formation of an objective and accurate cancer identification process in cancer diagnosis. In general, the input image size is much smaller than the histopathological image in most mainstream deep network architecture for classification tasks. A scanned histopathological image usually contains billions of pixels, needing image preprocessing into patches with appropriate sizes as the input for the deep network. A scanned histopathological image is divided into non-overlapping patches of the same size, and in this process, may generate thousands of image patches. Datasets will be manually annotated by pathologists at the patch level. The deep learning model will be trained on the above dataset to learn to extract the relevant features, guided by annotations. The more annotated the data, the higher classification performance of the deep learning model. For high-precision recognition models, deep learning methods are comparable to professional pathologists [45,61], but it is still difficult for most pathological models to form an effective comparison with pathologists. The biggest challenge of current classification problems is that training accurate classification models requires large datasets with physician annotations, which are often difficult to obtain. To deal with this problem, a commonly used practice at this stage is to expand the data by augmentation methods (e.g., color changes, random cropping, etc.). Collecting large datasets with annotations [62,63] and using pre-training methods such as transfer learning [64,65,66] and self-supervised learning [67,68] are effective methods to improve classification accuracy. In addition, the emergence of automated annotation procedures [69,70,71] provides an alternative to manual annotations by pathologists to reduce the requirements of professionals for data annotation.

Nuclei segmentation. Nuclear segmentation analyzes the features of cancer tissue and evaluates grade and prognosis. Nuclear features are significant in histopathology, as pathologists may diagnose the grading of patients or assess prognosis based on cell morphology and structure. For example, mitosis is closely related to breast cancer grading [72]. Nuclear density, size, and other indicators have a significant relationship with cancer prognosis [73]. The segmentation of nuclei from tissue images to extract nuclei features provides a premise for future analysis [74,75]. In a nuclei segmentation dataset, some representative regions of interest (ROIs) are selected from HI images by the pathologist, annotating the cell/nuclei contours and cell/nuclei types using annotation software. Typically, a single ROI contains hundreds of cells and several cell types, which takes great effort to annotate accurately segmentation ground truth. In nuclei segmentation methods, each pixel of the input image will be predicted as the probability of belonging to the cell/nuclei as output in the deep learning model. Pixel probabilities larger than a certain probability threshold are considered to belong to cell/nuclei region. One of the challenges in nuclei segmentation is the overlapping of nuclei and the gathered nuclei. Nuclei annotation is at the cell instance level, which is more difficult to annotate. Moreover, there are few publicly available datasets, leading to the segmentation method being difficult to expand between different datasets [76].

Transfer learning. Currently, there are several large datasets with complete annotation in some downstream tasks such as image classification [61], segmentation [77], and object detection [78,79]. In HI, datasets are usually annotated for a specific downstream task in a specific cancer or tissue [80,81,82], hindering study on other cancer types which other researchers are interested in. Transfer learning is a method to leverage knowledge from a source domain to improve the learning performance in a target domain [83]. Transfer learning can increase the performance and robustness of relatively small data in downstream analysis when it is difficult to obtain large datasets with annotation. In HI analysis, transfer knowledge gained from other large datasets to target the cancer domain can improve the performance of downstream tasks [54,84,85,86]. The difficulty of histopathological transfer learning is that most studies transfer the ImageNet dataset [61] to the medical data domain [87]. The ImageNet dataset contains natural objects, while the HI dataset contains repeated organizational structures and texture features. There are significant differences in the distribution of features and resolution between the two data formats. Based on the currently available large datasets of histopathology, suitable transfer learning methods could improve the performance of downstream tasks [87,88,89].

Self-supervised learning. Self-supervised learning is a method that obtains a representation of a dataset without annotation data, unsupervised to find common features among images from unlabeled datasets [90]. Self-supervised learning builds pretext tasks to learn a good representation of the inherent features in the training process, and fine-tunes downstream tasks using labeled data [91]. The pretext tasks of HI are generally constructed through contextual, multi-resolution, and semantic features [51,92], and the downstream tasks are analyzed by pre-training and fine-tuning. Effective features of HI are extracted using appropriate pretext tasks using prior information possessed in HI [93]. Self-supervised learning excavates image-learning semantic information from the image itself, which is the development direction of AI in medicine and of HI analysis in the future [94]. Learning how to design an appropriate pretext task for HI and obtaining general features are the key points in the application of self-supervised learning in HI.

In summary, AI has become an important research method for HI analysis. Breast cancer, prostate cancer, colorectal cancer, and lung cancer are the most studied cancer types. Classification and nucleus segmentation based on deep learning may continue to be research hotspots for a long time. Transfer learning and self-supervised learning are likely the development trends of future research. In general, this study provides some insights into the structure and trends in the field of AI HI analysis.

The authors thank the contribution of VOSviewer and CiteSpace bibliometric analysis tools.

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/ijerph191811597/s1↗. Table S1: Top 10 most productive countries of AI in HI from 2001 to 2021; Table S2: Top 10 most productive institutions of AI in HI from 2001 to 2021; Table S3: Top 10 most productive authors of AI in HI from 2001 to 2021; Table S4: Top 10 journals publishing research on AI in HI from 2001 to 2021; Table S5: Top 10 most frequently used keywords for methods and cancer types in publications of AI in HI.

Conceptualization, W.Z.; data curation, W.Z.; methodology, W.Z. and Z.D.; software, W.Z. and Z.D.; validation, Z.D. and Y.L.; writing—original draft preparation, W.Z.; writing—review and editing, Z.D., Y.L. and H.S.; supervision, H.X. and H.D.; funding acquisition, W.Z. All authors have read and agreed to the published version of the manuscript.

Not applicable.

Not applicable.

Not applicable.

The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

This research was funded by the Fundamental Research Funds for the Central Universities of Central South University grant number 1053320211619.