An AI-Driven Framework for Automated Image Preprocessing for Optimal Graph Extraction in StructuralGT

Accurate graph extraction from microscopy images is highly sensitive to the choice of preprocessing filters, a process that in StructuralGT currently relies on user intuition and extensive trial-and-error. We present a novel Artificial Intelligence (AI) module that integrates evolutionary optimization, deep learning, and reinforcement learning to automate this critical step. Our approach employs a genetic algorithm to explore trillions of possible image filter combinations, guided by a convolutional neural network (CNN) trained to discriminate between high- and low-quality graph representations. To further refine the search, we introduce a Markov Decision Process (MDP)-based reinforcement learning agent that adaptively steers the genetic algorithm toward filter sets that maximize graph quality. This pipeline ensures the generation of binary images that yield accurate and reliable graph skeletons for downstream analysis, including computation of betweenness centrality and other topological descriptors. By embedding this AI-driven module into StructuralGT, we provide researchers with a scalable, data-driven alternative to manual parameter tuning, enhancing reproducibility and accelerating the structural characterization of nanofibrous materials.