Interpretable Deep Learning of Endogenous Fluorophore Concentration and Entropy for Lung Tissue Classification

Fluorophore concentration measures and local histogram entropy describe different aspects of endogenous fluorophore maps. We investigate whether their combination improves discrimination between normal and cancerous lung tissue and which features contribute to predictions. Chemometric fluorescence microscopy (CFM) provides concentration maps of tryptophan, elastin, reduced nicotinamide adenine dinucleotide (NADH), and flavin adenine dinucleotide (FAD). A neural network based on generalized additive models with structured interactions (GAMI-Net) is used to compare concentration-only, entropy-only, and combined representations through explicit main effects and pairwise interactions. The mean areas under the receiver operating characteristic curve (ROC AUCs) across five-fold cross-validation were $95.9\% \pm 0.7\%$, $96.4\% \pm 1.6\%$, and $98.4\% \pm 0.7\%$, respectively; variability is reported as standard deviation (SD). Their aggregate differences are consistent with the hypothesis of complementary predictive information. FAD entropy, NADH concentration, and NADH entropy were the leading main effects in the combined model. The learned log-odds contributions include a U-shaped elastin-entropy relationship and nonmonotonic tryptophan relationships. By exposing how concentration and entropy enter predictions, this interpretable framework turns label-free fluorescence classification into specific, testable hypotheses about tissue metabolism and extracellular-matrix organization, motivating specimen-grouped and external validation.