Establishing a Highly Accurate Circulating Tumor Cell Image Recognition System for Human Lung Cancer by Pre-Training on Lung Cancer Cell Lines

Our plain-English summary of Establishing a Highly Accurate Circulating Tumor Cell Image Recognition System for Human Lung Cancer by Pre-Training on Lung Cancer Cell Lines, by H. Matsumiya, K. Terabayashi, Y. Kishi, et al (Cancers, 2025).

The study

• Authors: H. Matsumiya, K. Terabayashi, Y. Kishi, Y. Yoshino, M. Mori, M. Kanayama, R. Oyama, Y. Nemoto, N. Nishizawa, Y. Honda, T. Kuwata, M. Takenaka, Y. Chikaishi, K. Yoneda, K. Kuroda, T. Ohnaga, T. Sasaki, F. Tanaka
• Journal: Cancers
• Published: 2025;17(14)
• Read it: PubMed · Publisher · Free full text

Abstract

Background/objectives: Circulating tumor cells (CTCs) are important biomarkers for predicting prognosis and evaluating treatment efficacy in cancer. We developed the "CTC-Chip" system based on microfluidics, enabling highly sensitive CTC detection and prognostic assessment in lung cancer and malignant pleural mesothelioma. However, the final identification and enumeration of CTCs require manual intervention, which is time-consuming, prone to human error, and necessitates the involvement of experienced medical professionals. Medical image recognition using machine learning can reduce workload and improve automation. However, CTCs are rare in clinical samples, limiting the training data available to construct a robust CTC image recognition system. In this study, we established a highly accurate artificial intelligence-based CTC recognition system by pre-training convolutional neural networks using images from lung cancer cell lines.

Methods: We performed transfer learning of convolutional neural networks. Initially, the models were pre-trained using images obtained from lung cancer cell lines. The model's accuracy was improved by training with a limited number of clinical CTC images.

Results: Transfer learning significantly improved the CTC classification accuracy to an average of 99.51%, compared to 96.96% for a model trained solely on pre-trained cell lines (p < 0.05). This approach showed notable efficacy when clinical training images were limited, achieving statistically significant accuracy improvements with as few as 17 clinical CTC images (p < 0.05).

Conclusions: Overall, our findings demonstrate that pre-training with cancer cell lines enables rapid and highly accurate automated CTC recognition even with limited clinical data, significantly enhancing clinical applicability and potential utility across diverse cancer diagnostic workflows.

In plain English

This laboratory and computer study from Japan tested an artificial intelligence (AI) system that recognizes circulating tumor cells (cancer cells in the blood) in microscope images from the authors' "CTC-Chip" blood test. It did not study asbestos exposure or mesothelioma. The AI was first trained on images of three lung cancer cell lines and blood cells from two healthy people. It was then trained further, a step called transfer learning, on images from blood samples of 119 lung cancer patients, in which 201 tumor cells had been identified by eye. Average accuracy was 99.51% with transfer learning, compared with 96.96% for a model trained only on cell lines. That difference was statistically significant, and at least 17 clinical images were needed for a significant gain. Against a model trained on clinical images alone, the gain was statistically significant particularly with 19 or fewer training images. The authors concluded that pre-training on cell lines improved accuracy but that cell lines alone were insufficient. They noted the limited dataset size and the limited diversity of the samples.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Laboratory and computational (machine-learning) study that developed and tested an image-classification system for circulating tumor cells, comparing training approaches with and without pre-training on lung cancer cell lines
• Who was studied: Pre-training used images of three human lung cancer cell lines (PC9, NCI-H441, A549) and blood cells from two healthy individuals. Clinical images came from pre-treatment blood samples of 119 lung cancer patients diagnosed at the University of Occupational and Environmental Health, Japan, between 2014 and 2020. In these samples, 201 circulating tumor cells were visually identified, and 201 non-tumor cells were randomly selected to match.
• Exposure: No asbestos, fiber or occupational exposure was studied. The comparison was between three AI training approaches: pre-training on lung cancer cell line images followed by transfer learning with clinical tumor cell images; training on clinical images only; and pre-training on cell lines only.
• Main results:
  • Compared with a model trained on clinical images only, transfer learning gave higher classification accuracy at all training sample sizes up to 141 images. The difference was statistically significant particularly when 19 or fewer clinical training images were used (p less than 0.05).
  • At 141 clinical training images, average accuracy was 99.51% with transfer learning and 99.46% without it (clinical images only). The authors reported no statistically significant difference, indicating that pre-training did not degrade performance.
  • Average recognition accuracy was 96.96% for the model trained only on cell lines and 99.51% for the transfer-learning model, a statistically significant improvement of 2.55%. Minimum accuracy rose from 84.16% to 96.67%.
  • The improvement over pre-training alone was statistically significant when 17 or more clinical training images were used.
• The authors' conclusion: The authors concluded that pre-training on lung cancer cell lines improved classification accuracy for tumor cell images in clinical samples, even when clinical images were limited. They added that training on cell lines alone was insufficient and that clinical images were also needed.
• Limitations the authors state:
  • Only lung cancer cell lines were used. Whether the approach works for other cancer types and cell lines needs further validation.
  • Clinical samples were collected under limited conditions and may not reflect real-world diversity in patient backgrounds or collection protocols.
  • The dataset was limited in size, and a larger, multi-institutional dataset is needed.
  • The study evaluated only image classification, not the whole detection process (capture, staining and data acquisition).
  • Other deep learning architectures were not compared.

Topics

Medical terms in this study, explained in our Lexicon: Lung Cancer, Biomarker, Prognosis, Pleural Mesothelioma.

Related studies

• Detection of circulating tumor cells with a novel microfluidic system in malignant pleural mesothelioma — K. Yoneda, T. Kuwata, et al. (2019)
• Novel circulating tumor cell-detection chip combining conventional podoplanin and EGFR antibodies for all histological malignant pleural mesothelioma — M. Kanayama, R. Oyama, et al. (2021)
• Detection of Circulating Tumor Cells (CTCs) in Malignant Pleural Mesothelioma (MPM) with the "Universal" CTC-Chip and An Anti-Podoplanin Antibody NZ-1.2 — T. Kuwata, K. Yoneda, et al. (2020)
• EpCAM-independent capture of circulating tumor cells with a 'universal CTC-chip' — Y. Chikaishi, K. Yoneda, et al. (2017)
• Impact of CD73/CD155 Expression on Survival in Resected Epithelial Pleural Mesothelioma: A Retrospective Single-center Study — Y. Nabe, Y. Fujita, et al. (2026)
• Circulating tumor cells (CTCs) in malignant pleural mesothelioma (MPM) — K. Yoneda, F. Tanaka, et al. (2014)

Browse more research in Cancers, or return to medical research.