Deep learning-based cell segmentation for rapid optical cytopathology of thyroid cancer

Area of Science:

• Biomedical Imaging
• Computational Biology
• Oncology

Background:

• Fluorescence polarization (Fpol) imaging of methylene blue (MB) shows potential for quantitative thyroid cancer detection.
• Clinical application of MB Fpol technology is hindered by lengthy data analysis times.

Purpose of the Study:

• To develop and evaluate a deep learning-based automated cell segmentation method for MB Fpol imaging.
• To reduce data analysis time and assess the accuracy of automated versus manual processing for thyroid cancer diagnosis.

Main Methods:

• A 2D U-Net convolutional neural network was trained and tested on diverse human thyroid cell images.
• Automated (AU) and manual (MA) data processing methods were compared for cell count, segmented area, and Fpol values.

Main Results:

• The automated model segmented 15.8% more cells than manual analysis.
• Differences in segmented cell areas ranged from -55.2% to +31.0% between AU and MA.
• Fpol value differences ranged from -20.7% to +10.7%, with no statistically significant differences observed between AU and MA.
• Analysis time was reduced from one hour (MA) to 10 seconds (AU).

Conclusions:

• Deep learning-based automated cell segmentation is effective for MB Fpol imaging in thyroid cancer detection.
• The automated method significantly reduces analysis time while maintaining diagnostic accuracy.
• This approach makes quantitative fluorescence polarization-based diagnosis clinically feasible.