Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2023 Aug 26;10(1):570.
doi: 10.1038/s41597-023-02482-8.

Label-free tumor cells classification using deep learning and high-content imaging

Chawan Piansaddhayanon #  1   2   3 , Chonnuttida Koracharkornradt #  2 , Napat Laosaengpha  1   2 , Qingyi Tao  4 , Praewphan Ingrungruanglert  5 , Nipan Israsena  6   7 , Ekapol Chuangsuwanich  8   9 , Sira Sriswasdi  10   11
Affiliations

Label-free tumor cells classification using deep learning and high-content imaging

Chawan Piansaddhayanon et al. Sci Data. .

Abstract

Many studies have shown that cellular morphology can be used to distinguish spiked-in tumor cells in blood sample background. However, most validation experiments included only homogeneous cell lines and inadequately captured the broad morphological heterogeneity of cancer cells. Furthermore, normal, non-blood cells could be erroneously classified as cancer because their morphology differ from blood cells. Here, we constructed a dataset of microscopic images of organoid-derived cancer and normal cell with diverse morphology and developed a proof-of-concept deep learning model that can distinguish cancer cells from normal cells within an unlabeled microscopy image. In total, more than 75,000 organoid-drived cells from 3 cholangiocarcinoma patients were collected. The model achieved an area under the receiver operating characteristics curve (AUROC) of 0.78 and can generalize to cell images from an unseen patient. These resources serve as a foundation for an automated, robust platform for circulating tumor cell detection.

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing interests.

Figures

Fig. 1
Fig. 1
Examples of preprocessed and annotated brightfield and fluorescence image for human annotator. Box colors indicate the object classes (red for cancer cells, green for normal cells, and blue for unknown cells that exhibited neither signals).
Fig. 2
Fig. 2
Examples of annotated cells from each class.
Fig. 3
Fig. 3
The annotation process for the training and validation sets. First, a small subset (30 images) was fully manually annotated. Then, the initial cell detection and classification model was trained to generate pseudolabels for all unannotated images. The pseudo-generated bounding boxes were then filtered using Non-Maximum Suppression (NMS) to remove highly overlapping boxes. These pseudolabel annotation were then refined by the experts to obtain the final annotation used for training and validation. Note that every step in this annotation process used fluorescence images as guidance.
Fig. 4
Fig. 4
The annotation process for the test set. Two annotators were separately tasked to annotate all cancer cells inside each brightfield image with paired fluorescence image as guidance. Results from the two annotators were combined and used as the final annotation.
Fig. 5
Fig. 5
The index of our proposed dataset.
Fig. 6
Fig. 6
The main pipeline for cancer cell detection consists of two stages, detection and classification, each being a deep artificial neural network. The detector proposes possible cancer cells which are then re-examined by the classifier to refine the confidence scores. Finally, Non-Maximum Suppression (NMS) is performed to remove highly overlapping bounding boxes.
Fig. 7
Fig. 7
Normalized confusion matrix of the cell-level evaluation on the validation split.
Fig. 8
Fig. 8
2D embeddings of cells from different classes in the dataset. The embeddings were calculated using UMAP from the feature map at the last layer before the last global average pooling in the network. (a) Embeddings from the model trained with brightfield images and all fluorescence signals. (b) Embeddings from the model trained using only brightfield images. (c) Embeddings from the model trained with brightfield images and Hoeschst signal.
Fig. 9
Fig. 9
Example of image-level predictions (red boxes) and their confidence on the test set under the Brightfield + Fluorescence setting. Despite having the fluorescence signal as guidance, the model still outputted oversized bounding boxes and could not distinguish individual cells in areas with high cell density.
Fig. 10
Fig. 10
2D embeddings of cells from different patients in the dataset. The embeddings were calculated using UMAP from the feature map at the last layer before the last global average pooling in the network. (a) Embeddings from the model trained with brightfield images and all fluorescence signals. (b) Embeddings from the model trained using only brightfield images. (c) Embeddings from the model trained with brightfield images and Hoeschst signal.
Fig. 11
Fig. 11
Impact of the training set size on the AUROC and F1 performances of cancer cell classification. Performances of the classifier were measured on the validation set. With full florescence signals as input, the model readily learned to identify cancer cells even with only a small data subset (green curve). In other settings, performances increased linearly as the data grew exponentially.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

References

    1. Rawal, S., Yang, Y.-P., Cote, R. & Agarwal, A. Identification and quantitation of circulating tumor cells. Annual Review of Analytical Chemistry10, 321–343, 10.1146/annurev-anchem-061516-045405. PMID: 28301753 (2017). - PubMed
    1. Ming, Y. et al. Circulating tumor cells: From theory to nanotechnology-based detection. Frontiers in Pharmacology8 (2017). - PMC - PubMed
    1. Bankó, P. et al. Technologies for circulating tumor cell separation from whole blood. Journal of Hematology & Oncology12, 10.1186/s13045-019-0735-4 (2019). - PMC - PubMed
    1. Satelli, A., Brownlee, Z., Mitra, A., Meng, Q. & Li, S. Circulating tumor cell enumeration with a combination of epithelial cell adhesion molecule- and cell-surface vimentin-based methods for monitoring breast cancer therapeutic response. Clinical chemistry61, 10.1373/clinchem.2014.228122 (2014). - PMC - PubMed
    1. Xu, Y. et al. Circulating tumor cell detection: A direct comparison between negative and unbiased enrichment in lung cancer. Oncology Letters13, 10.3892/ol.2017.6046 (2017). - PMC - PubMed