Systematic Literature Review of Computer Vision-Aided Cytology - A Review of Classic Computer Vision and Deep Learning-Based Approaches published between January/2016 - March/2020

Access the paper here.

Abstract

Cytology is a low-cost and non-invasive technique where cells are harvested from tissues by aspiration or scraping that is used to diagnose a broad range of pathologies. Although this procedure is performed by clinical consultants extensively trained for that, it is a time consuming and repetitive process where most of the diagnostic criteria are vulnerable to human interpretation and failure. Hence, the use of computer technologies can reduce the chances of misdiagnoses, shorten the time required for the analysis and help on early cancer diagnosis. In order to identify the state-of-art of computer vision techniques applied on cytology images, we conducted a Systematic Literature Review based on the protocols elaborated by Barbara Kitchenham, searching for approaches to cell segmentation, detection, and classification using computer vision on cytology slides images and analyzing papers published in the last years. The initial search was made on March 6th, and resulted in articles that we analyzed based on the inclusion/exclusion criteria, resulting in papers that we used to evaluate the tendencies and main problems present in this research area, highlighting the computer vision methods, staining techniques, evaluation metrics, and the availability of the used datasets. As a result, we found out that the most used methods in the analyzed works are deep learning-based (papers), especially Convolutional Neural Networks (papers), while fewer works use classic computational vision only methods (papers). The most recurrent metric used was the accuracy for classification (papers) and the Dice Similarity Coefficient for segmen-tation and object detection (papers). Regarding staining techniques, Papanicolaou was the most used one (papers), followed by H&E (papers) and Feulgen (papers). Among the datasets used in the papers, 8 of them are publicly available, with ISBI and ISBI being the most used ones. Analyzing the papers found on this review, we concluded that there is a lack of high-quality datasets for some types of stain, most of the works are not ready to be applied on a daily clinical routine, and there is a tendency to adopt deep learning methods in the future.

Bibtex

@techreport{SLRCytoINCoD2020,
    doi = {10.13140/RG.2.2.13304.67840},
    url = {http://rgdoi.net/10.13140/RG.2.2.13304.67840},
    author = {Jo{\~{a}}o Gustavo Atkinson Amorim and Allan Cerentini and Luiz Antonio Buschetto Macarini and  Andr{\'{e}} Vict{\'{o}}ria Matias and Aldo von Wangenheim},
    language = {en},
    title = {Systematic Literature Review of Computer Vision-Aided Cytology - A Review of Classic Computer Vision and Deep Learning-Based Approaches published between January/2016 - March/2020},
    publisher = {Federal University of Santa Catarina},
    year = {2020}
}