# From Tissue Architecture To Genetic Signature: Artificial intelligence-based Analysis of Reticulin Framework and Clinical Variables Predicts Molecular Cluster in Paragangliomas

## TL;DR

This study uses artificial intelligence to analyze reticulin patterns in paragangliomas and predict genetic clusters, linking tissue architecture with molecular classification.

## Contribution

A novel deep learning model is introduced to predict germline cluster 1 genotype using reticulin framework analysis and clinical variables.

## Key findings

- AI-based models achieved high accuracy (AUC 0.981-0.990) in predicting germline cluster 1 genotype.
- Germline cluster 1 tumors showed distinct architectural features like intact reticulin and very small nests.
- Combining AI-derived metrics with clinical data improves pre-test genetic prediction reliability.

## Abstract

Paragangliomas (PGLs), encompassing pheochromocytomas and extra-adrenal paragangliomas, are genetically heterogeneous non-epithelial neuroendocrine neoplasms that segregate into molecular clusters with distinct biological and clinical behavior. Architectural correlates of genotype have not been systematically investigated. This study aimed to assess the reticulin framework as a potential morphologic structural surrogate of molecular background and to introduce a novel deep learning model for its spatially resolved quantitative analysis and genotype prediction. A total of 104 adrenal and extra-adrenal PGLs with complete clinical, pathological, and genetic data were retrospectively analyzed. Reticulin stain was evaluated qualitatively and quantitatively using a supervised convolutional neural network trained on expert-annotated reticulin-stained whole-slide images (WSIs) to map and quantify areas of intact framework and very small nest patterns. Two bias-reduced logistic regression models (Firth’s method) were developed to predict germline cluster 1 genotype, each combining clinical variables (age, tumor size, extra-adrenal presentation) with one artificial intelligence (AI)-derived morphometric feature-percentage of intact framework (Model-INTACT) or very small nests (Model-VSN). PGLs harboring germline cluster 1 variants occurred at a younger age, were larger, more frequently extra-adrenal, and showed significant enrichment of intact reticulin and very small nest patterns compared with cases harboring germline cluster 2 variants and sporadic cases. The supervised AI model accurately mapped and quantified these architectural features across the WSIs. Predictive models integrating AI-derived morphometrics with clinical variables achieved excellent discrimination for germline cluster 1 genotype (AUC 0.981 for Model-INTACT; AUC 0.990 for Model-VSN). Preservation of the reticulin framework, particularly with very small nests, represents a histoarchitectural correlate of pseudohypoxic PGLs. Integration of AI-based morphometric descriptors with clinical parameters enables reliable pre-test prediction of germline cluster 1 genotype, bridging conventional histopathology and molecular classification.

The online version contains supplementary material available at 10.1007/s12022-026-09904-4.