A dual-branch deep learning network for circulating tumor cells classification.

Abstract

Background: Circulating tumor cells (CTCs) in peripheral blood are crucial for prognosis, treatment response, disease monitoring, and personalized therapy. However, identifying CTCs remains challenging due to their scarcity and heterogeneity, even with advanced deep learning models.

Methods: This study introduces an innovative hybrid framework combining a dual-branch network with traditional image processing techniques and automated CTC identification. By incorporating image and fluorescence attributes, the framework enhances feature representation robustness. Performance was evaluated using accuracy, precision, and recall metrics and comparisons with pathologists' manual counting.

Results: The framework achieved 97.05% accuracy in distinguishing CTCs from non-CTCs, with performance closely matching pathologists' manual counting in survival prediction. The dual-branch network improved efficiency by leveraging segmentation algorithms, surpassing conventional methods. Clinical trials confirmed its practicality for direct clinical use.

Conclusions: The proposed framework enhances CTC identification accuracy and efficiency, demonstrating strong clinical applicability. Its output results can be directly utilized for prognosis without manual intervention, offering significant potential for personalized therapy.