Multi-Parameter Deep Learning Combined With Fluorescence Lifetime Imaging Microscopy for Non-Invasive and Label-Free...

This study presents a noninvasive and label-free endometrial cancer screening approach combining fluorescence lifetime imaging microscopy (FLIM) with a multi-parameter deep learning algorithm. The autofluorescence signals of NAD(P)H in cervical exfoliated cells of participants (n = 71) were detected by FLIM. Two key metabolic parameters-mean fluorescence lifetime (t_m) and protein-bound fraction (a₂)-were extracted using FLIM data, which reflect malignancy-associated biochemical changes. Single-parameter models based on individual parameters (t_m or a₂) and a multi-parameter model integrating both parameters were constructed. The results showed that single-parameter models suffered from sensitivity-specificity imbalance while the multi-parameter model leveraging feature complementarity significantly improved predictive performance-achieving 100% sensitivity and 92% specificity in external testing, with an AUC of 0.92, representing an improvement of 0.17-0.31 over single-parameter models. These findings suggest that combining multi-parameter deep learning strategies with FLIM holds strong potential for risk prediction of endometrial cancer, offering a new approach for noninvasive clinical screening.