Analysis of circulating tumor cells (CTCs) provides exclusive benefits compared with other liquid biopsy techniques due to their advanced clinical applications including single-cell multiomics (DNA, RNA, proteins, and metabolites) analysis and the ability to generate CTCs-derived xenograft models. Current CTC analysis primarily relies on labor-intensive and human-biased manual CTC enumeration or semiautomated work, hindering clinical adaptation of CTC analysis into practice. Beyond CTC enumeration, recent studies focused on detection and characterized features of circulating tumor microemboli (CTM) to improve cancer monitoring and diagnosis, but the added complexity of CTM makes CTC analysis even more challenging. Development of an automated, unbiased, and specialized computation method for CTC analysis remains a clinical unmet need. We developed a deep learning-based framework by applying computer vision to efficiently automate CTC enumeration and characterization with high throughput and accuracy. Fluorescence microscopy CTCs images (CK+/CD45-/DAPI+) have been prepared from blood samples of non-small cell lung cancer patients by the “CMx” CTC capture platform. This study applied an “active learning” model that only requires a small upfront training dataset while being able to accumulate more training data to make improvements with minimal annotation efforts. We have collected 20 images annotated by subject matter experts to train the initial model. Using an iterative process where feedback from the subject matter experts was used to influence the model, we achieved continuous improvements using adaptive annotations. After the model had been well trained and validated by 4 extra images, we used 18 new test images to benchmark the model performance. Preliminary results demonstrated that our AI model outperformed current semiautomated methods having higher sensitivity and reduced time for CTC enumeration in a lung cancer CTC sample analysis. In 18 test samples, the model predicted 34% more total CTCs than the current method (1,775 versus 1,328 counts). By applying the model to the 24 training samples, the model also recovered 45% more total CTCs absent from the original human annotation (2,507 versus 1,732 counts). In some samples, enumeration assisted by the model saved 90% of time over the current method (<20 min vs ~4 hrs.) for CTC enumeration, which would result in an enormous advantage if the datasets being analyzed included a large number of samples. Furthermore, our AI model was able to characterize features of CTM, including CTC clusters and CTC-associated immune cells, in addition to enumerating the samples. The AI model developed in this study is a promising method for CTC analysis that significantly improved throughput, accuracy, and reproducibility, leading to an approach of CTC analysis better suited for clinical adoption. Development of AI-driven CTC/CTM analysis provides additional resources and advancements to the field of liquid biopsy for the diagnosis and treatment of disease.

Citation Format: Wei-Lei Yang, Chi-Bin Li, Jia-Yang Chen, Samuel Chen, Ten-Jen Liu, Ying-Chih Chang. Artificial intelligence assists automation and high performance of circulating tumor cells enumeration and circulating tumor microemboli characterization in fluorescence microscopy images [abstract]. In: Proceedings of the AACR Special Conference on Advances in Liquid Biopsies; Jan 13-16, 2020; Miami, FL. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(11_Suppl):Abstract nr B14.