<p>The potential of electrical capacitance tomography in multiphase flow measurements is constrained by low-quality images. To alleviate this challenge, this study models the imaging task as a bilevel optimization framework. This new model integrates principles of supervised learning and optimization, addresses the inaccuracies in both the reconstruction model and measurement data, integrates measurement principles with machine learning, enables adaptive learning of model parameters and prior images, achieves multisource information fusion, mitigates the ill-posedness, and enhances the automation and robustness of the model. A new bilevel multimodal extreme learning machine is developed to improve the inference accuracy of prior images. A novel optimizer is proposed to solve the bilevel optimization imaging model. This algorithm reduces computational complexity by converting the original bilevel optimization problem into a single-level surrogate optimization problem, which integrates principles of learning and optimization, allowing performance to be improved through learning during the optimization process. Qualitative and quantitative evaluation results indicate that, compared to popular imaging algorithms, the novel method enhances imaging quality and improves noise immunity. This study provides new insights and methodologies to maximize the potential of this measurement technology.</p>