Autofluorescence intensity patterns encode α/β cell identity in human islets

Technical Summary: Autofluorescence Intensity Patterns Encode α/β Cell Identity in Human Islets

Problem Statement
Reliable identification of pancreatic α- and β-cells within intact human islets is critical for studying islet physiology and diabetes pathogenesis. Current gold-standard methods, such as immunohistochemistry, require cell fixation and are destructive, precluding live-cell analysis. Alternative live-cell strategies often rely on genetically encoded reporters, which necessitate exogenous manipulation and may alter native cellular states. While Two-Photon Fluorescence Lifetime Imaging Microscopy (2PM-FLIM) has emerged as a powerful non-invasive tool, it requires specialized instrumentation and complex acquisition protocols. Previous multimodal machine-learning approaches combining autofluorescence intensity with fluorescence-lifetime parameters achieved moderate discrimination (AUC = 0.86), but it remained unresolved whether spatial intensity information alone—without lifetime data—was sufficient for robust cell-type discrimination.

Methodology
The study presents a retrospective computational re-analysis of a previously published experimental dataset of living human islets. The workflow isolates the spatial intensity component of autofluorescence images, deliberately discarding fluorescence lifetime (FLIM) information.

• Data Curation: The dataset comprised 1,271 single-cell images (634 β-cells and 605 α-cells) derived from two-photon microscopy (740 nm excitation, 420–460 nm emission). Cells were manually segmented, and overlapping cells (>10% area overlap) were excluded. Images were standardized to a 150×150 pixel canvas.
• Feature Extraction: A 60-dimensional feature vector was extracted from each image, consisting of:
  • Morphological Features (40 dimensions): Including 22 classical shape descriptors, 7 log-transformed Hu moments (rotation-invariant), and 11 Fourier descriptors.
  • Texture Features (20 dimensions): Quantified using Rotation-Invariant Uniform 2 (RIU2) Local Ternary Pattern (LTP) histograms to capture fine-scale intracellular intensity heterogeneity.
• Machine Learning Pipeline: The dataset was split into 80% training and 20% held-out test sets. Nine classifiers (including Logistic Regression, SVMs, Random Forest, LightGBM, and XGBoost) were optimized using Bayesian hyperparameter tuning within a 5-fold stratified cross-validation scheme. Data augmentation (rotation and reflection) was applied to ensure geometric invariance. Ensemble strategies (voting and stacking) were also evaluated.
• Interpretability: The final model was analyzed using Permutation Feature Importance, Partial Dependence Plots (PDPs), and Shapley Additive Explanations (SHAP) to identify the biological drivers of classification.

Key Results

• Classification Performance: The proposed intensity-only framework achieved a mean Area Under the Receiver Operating Characteristic Curve (AUC) of 0.92 ± 0.02 for the 'All models Voting' ensemble, significantly outperforming the previous multimodal benchmark of 0.86. The Light Gradient Boosting Machine (LGBM) model, selected for parsimony, performed statistically comparably to the complex ensembles.
• Robustness: The model demonstrated high stability under stress tests involving random image rotations and reflections, with performance drops remaining below one percentage point and no statistically significant differences observed.
• Feature Importance: Interpretability analyses revealed that predictions were driven predominantly by fine-scale intracellular intensity organization (specifically LTP texture features) rather than global morphology.
  • LTP Upper Pattern 7 emerged as the dominant predictor. Its value showed a strong, non-linear, sigmoidal relationship with the probability of a cell being a β-cell.
  • Secondary features included morphological descriptors like Max Defect Depth and Inertia Eigenvalue 1, but their contribution was markedly lower than that of texture features.
• Biological Correlation: The reliance on texture features aligns with the biological observation that β-cells possess a higher abundance of lipofuscin-rich granules compared to α-cells. These granules create heterogeneous intracellular intensity patterns that LTP descriptors effectively capture.

Significance and Claims
The paper claims that endogenous autofluorescence intensity patterns contain sufficient structural information for reliable α/β cell discrimination without the need for fluorescence lifetime information. This finding supports a simpler, biologically interpretable route to prospective label-free cell identification.

The authors emphasize that while the method simplifies the acquisition workflow by removing the need for FLIM hardware and processing, the results are specific to the acquisition conditions of the dataset used (two-photon microscopy). The study does not claim immediate cross-platform applicability to different optical configurations or intensity-only datasets acquired independently; prospective validation on such datasets is required to establish transferability. Furthermore, the association between the identified texture patterns and lipofuscin granules remains an interpretation of image-derived features, requiring future correlative imaging or targeted experimental validation for molecular confirmation.

In summary, the work demonstrates that spatial intensity organization, rather than just global morphology or metabolic lifetime data, encodes cell identity in human islets, offering a potential pathway for more accessible, label-free analysis of islet physiology.