Deep Learning-based Image Cytometry Using a Bit-pattern Kernel-filtering Algorithm to Avoid Multi-counted Cell...

Area of Science:

• Pathology
• Computational Biology
• Oncology

Background:

• Digitization of tissue slides and deep learning image analysis enhance pathological diagnosis and patient management.
• Deep learning-based image cytometry (DL-IC) is crucial for precise cell identification and counting in digital pathology.
• Accurate cell determination is fundamental for the effective application of DL-IC techniques.

Purpose of the Study:

• To evaluate the performance of a novel DL-IC system, Cu-Cyto, in cell identification.
• To assess the impact of immunohistochemical staining on DL-IC performance.
• To determine the potential of DL-IC in advancing precision oncology.

Main Methods:

• Development of Cu-Cyto, a DL-IC utilizing a bit-pattern kernel-filtering algorithm to prevent cell miscounting.
• Evaluation of Cu-Cyto's performance on tumor tissue slide images with immunohistochemical staining (IHC).
• Assessment of three Cu-Cyto versions across different learning stages.

Main Results:

• Early in training, Cu-Cyto showed higher F1 scores for immunostained CD8+ T cells (0.343) compared to non-immunostained cells (adenocarcinoma: 0.040, lymphocytes: 0.002).
• Performance improved for all cell types as training and validation progressed.
• In the final learning stage, F1 scores reached 0.589 for adenocarcinoma cells, 0.889 for lymphocytes, and 0.911 for CD8+ T cells.

Conclusions:

• Cu-Cyto demonstrates effective cell determination capabilities.
• Immunohistochemical staining significantly enhances DL-IC learning efficiency, particularly in early stages.
• Continued learning is expected to further improve Cu-Cyto's performance, supporting precision oncology implementation.