Deep learning of circulating tumour cells

• Published in: Nature machine intelligence Vol. 2; no. 2; pp. 124 - 133
• Main Authors: Zeune, Leonie L., Boink, Yoeri E., van Dalum, Guus, Nanou, Afroditi, de Wit, Sanne, Andree, Kiki C., Swennenhuis, Joost F., van Gils, Stephan A., Terstappen, Leon W.M.M., Brune, Christoph
• Format: Journal Article
• Language: English
• Published: Basingstoke Nature Publishing Group 01.02.2020
• Subjects: Artificial intelligence; Artificial neural networks; Automation; Biomarkers; Blood circulation; Cancer; Classification; Cytokeratin; Datasets; Deep learning; Fluorescence; Image analysis; Machine learning; Medical imaging; Medical prognosis; Tumors; Visualization
• ISSN: 2522-5839; 2522-5839
• DOI: 10.1038/s42256-020-0153-x

Circulating tumour cells (CTCs) found in the blood of cancer patients are a promising biomarker in precision medicine. However, their use is currently hindered by their low frequency, tedious manual scoring and extensive cell heterogeneities. Those challenges limit the effectiveness of classical machine-learning methods for automated CTC analysis. Here, we combine autoencoding convolutional neural networks with advanced visualization techniques. This provides a very informative view on the data that opens the way for new biomedical research questions. We unravel hidden information in the raw image data of fluorescent images of blood samples enriched for CTCs. Our network classifies fluorescent images of single cells in five different classes with an accuracy, sensitivity and specificity of over 96%, and the obtained CTC counts predict the overall survival of cancer patients as well as state-of-the-art manual counts. Moreover, our network excelled in identifying different important subclasses of objects. Deep learning was faster and superior to classical image analysis approaches and enabled the identification of new biological phenomena.Counting different types of circulating tumour cells can give valuable information on the severity of the disease and on whether treatments are effective for a specific patient. In this work, the authors show that their method based on autoencoders can identify and count cells more accurately and faster than human experts.