Viscosity changes in cervical exfoliated cells: A non-invasive deep learning approach for endometrial cancer detection

This study investigates a novel non-invasive screening method for endometrial cancer (EC) based on the theory of field cancerization, utilizing cervical exfoliated cells from a cohort of 96 participants across three hospital branches. Cells were stained with a viscosity-sensitive fluorescent probe, and fluorescence lifetime imaging microscopy (FLIM) was employed to generate a substantial dataset of images. Two deep learning models were developed to predict EC based on these images. The model relying solely on cellular morphology (Model A) demonstrated 98.1 % training accuracy with suboptimal diagnostic performance (AUC = 0.79). In contrast, the advanced model incorporating both cellular morphology and intracellular viscosity information (Model B) achieved superior performance with 99.3 % training accuracy and significantly improved diagnostic capability (AUC = 0.90). External validation of Model B showed complete discrimination between EC and non-EC cases with sensitivity of 100 % (95 % CI:61.0–100 %) and specificity of 100 % (95 % CI:89.6–100 %). The findings underscore the potential of combining morphological and intracellular microenvironment viscosity data to enhance the accuracy of EC detection, offering a promising advance in early cancer screening.