Deep learning model utilizing fluorescence lifetime imaging microscopy and confidence learning for predicting endometrial cancer risk

Abstract

This study introduces a novel diagnostic approach for the prediction of endometrial cancer risk, which involves minimally invasive exfoliation cytology specimens from 63 participants, staining with the viscosity-sensitive fluorescent probe DCVJ, and detection by fluorescence lifetime imaging microscopy (FLIM). Initially, DCVJ-stained tissue sections from cancerous and normal regions were utilized to confirm the elevated viscosity in cancerous lesions. Subsequently, DCVJ-stained aspirated cell samples were detected by FLIM to capture viscosity characteristics and cellular morphology. A total of 2348 FLIM images collected from 42 individuals, were utilized for training purposes. A deep learning model, incorporating confidence learning, was developed and demonstrated a sensitivity of 84.6% and specificity of 75.0% in predicting endometrial cancer risk. This study highlights the potential of the deep learning-FLIM model for the prediction of endometrial cancer risk.