Three-dimensional inversion of logging-while-drilling (LWD) data serves as a key technology for boundary detection and reservoir characterization under complex geological conditions. Currently, inversion methods based on deep neural networks are prone to overfitting, often yielding unreliable results when training samples are limited. To address this issue, this paper proposes a transfer learning-based electromagnetic inversion approach for LWD, adopting a strategy of 2D pre-training followed by 3D fine-tuning. This strategy enables efficient transfer of electromagnetic features learned from 2D models to 3D networks, thereby reducing the dependency of 3D inversion on large-scale datasets. Furthermore, a Bayesian Neural Network (BNN) is integrated to construct a unified 3D inversion framework, achieving simultaneous prediction of resistivity distribution and quantification of uncertainty. Numerical experiments demonstrate that the proposed method achieves fast and stable convergence even with a limited number of 3D training samples, while maintaining high inversion accuracy under varying noise levels. In addition, uncertainty estimation effectively reveals the impact of inadequate prior knowledge and noise interference on the inversion results. By ensuring inversion accuracy and significantly improving computational efficiency, the proposed method provides robust technical support for electromagnetic LWD inversion in complex geological settings.