Neural Network-Based System for Rapid Diagnostic Decision Support Using FLIM Data: Segmentation and Classification in Liver, Skin, and Pancreatic Tissues

Fluorescence lifetime imaging microscopy (FLIM) is a powerful tool for analyzing metabolic changes by measuring the fluorescence lifetime of endogenous fluorophores such as the reduced form of nicotinamide adenine dinucleotide (phosphate) (NAD(P)H), enabling differentiation between normal and pathological states in diverse tissues. However, traditional manual FLIM data processing suffers from several drawbacks, including subjective selection of regions of interest, variability in analysis parameters, time-consuming procedures, and inherent human bias. To address these limitations, we developed an automated pipeline that leverages multiple pre-trained foundational models—SAM 2, CellposeSAM, LACSS, and CellSAM—for rapid, intensity-based segmentation and subsequent classification of the FLIM parameters in liver, pancreas, and skin tissues. Our comparative analysis identified the most suitable models for each tissue type and, crucially, enabled accurate discrimination between normal and pathological conditions. SAM 2 and LACSS demonstrated stable segmentation performance across both healthy and diseased tissues, though with distinct strengths: SAM 2 excelled at segmenting large structures (e.g., entire pancreatic islets of Langerhans, or very large cells), whereas LACSS was more effective for small structures. In contrast, CellposeSAM and CellSAM yielded less consistent segmentation overall but achieved notably high efficiency in specific cases. These findings establish a foundation for the rapid, automated discrimination of normal versus pathological tissues, offering a robust complement to existing clinical methods and to FLIM workflows for research.