Abstract: Internal corrosion is a major threat to the safety of natural gas pipelines, with defect geometry—depth, length, and width—playing a critical role in structural integrity assessments. While corrosion depth prediction has been widely studied, systematic probabilistic modeling of defect length and width remains limited. This study develops a hierarchical Bayesian-Markov Chain Monte Carlo (HB-MCMC) framework to jointly predict corrosion defect dimensions from in-line inspection (ILI) data. The framework integrates non-centered parameterization and adaptive sampling to improve inference efficiency and employs a hierarchical dynamic thresholding procedure for robust data preprocessing and outlier filtering. Field data from two transmission pipelines in Southwest China, comprising 1845 defect records, are analyzed. Results demonstrate that defect length and width both increase with depth, with width exhibiting stronger sensitivity. Model diagnostics confirm convergence and reliable uncertainty quantification. To further explore underlying mechanisms, OLGA multiphase flow simulations are combined with statistical predictions, providing flow-parameter profiles along the pipelines and enabling correlation analysis between local hydrodynamics and defect geometry. The proposed framework not only enhances predictive capability for defect length and width but also provides new insights into flow-corrosion interactions under real operating conditions, offering a reproducible and data-driven tool for corrosion assessment.