Abstract: The investigation into the evolution mechanisms of fractures in high-stress-sensitive coal under external loading and their controlling effects on permeability is crucial for understanding gas migration in deep coal seams. To address the limitations of traditional image analysis in characterizing fractures in coal CT images, this study employs a deep learning algorithm to automatically segment fractures. Based on in-situ CT experimental data under confining pressures ranging from 0 to 25 MPa, the three-dimensional structure and topological model of the fracture network were reconstructed, and the evolution of structural parameters was quantified. In terms of theoretical modeling, unlike conventional methods that average fracture properties, a penny-shaped fracture assumption was introduced to equivalently characterize the progressive closure of fractures under stress. A critical stress model was subsequently developed to reflect the fundamental closure stress conditions of the fracture system. Based on this theory, a permeability evolution model applicable to high-stress conditions was further derived. The results show that the deep learning model achieved a prediction accuracy of 84.14% on out-of-distribution images, significantly outperforming the traditional method (40.96%), thereby effectively improving fracture extraction accuracy. As the confining pressure increased from 0 MPa to 6 MPa, the average fracture aperture decreased from 136.96 μm to 75.29 μm, the average tortuosity increased from 2.87 to 3.32, and the average coordination number—a key indicator of connectivity—dropped from 2.45 to 1.21. At a confining pressure of 25 MPa, the fracture porosity decreased by an average of 98.99%, indicating near-complete closure of the fractures. Based on hydrostatic compression test data and a trial-and-error iterative algorithm, the critical closure stresses were determined to be 24.16 MPa and 23.86 MPa, with corresponding permeabilities as low as 8.78×10⁻⁷ μm²and 4.20×10⁻⁶ μm², respectively. The converted diffusion coefficients matched the magnitude measured in coal particle desorption experiments, indirectly confirming that the critical-stress condition exerts a sealing effect on seepage pathways. Compared to traditional models, the proposed permeability evolution model, which accounts for both fracture closure and elastic compression, significantly improved prediction accuracy—reducing the average error by 51.14%, particularly under high-stress conditions above 20 MPa.