<p>Electrical capacitance tomography suffers from low accuracy images, in which imaging method plays a pivotal role in addressing this challenge. To maximize the potential and effectiveness of the measurement technique while addressing the challenges inherent in imaging tasks, we conceptualize the imaging problem as a novel learnable bilevel fractional optimization problem. This approach, comprising upper and lower-level optimization problems, achieves adaptive learning of model parameter and image prior, synergizes machine learning with measurement principles, alleviates the ill-posed property of the imaging model, and improves the model automation. A new optimizer that integrates the particle swarm optimization algorithm and the custom-designed fractional optimization solver is proposed to solve the learnable bilevel fractional optimization imaging model. A new bilevel graph convolutional extreme learning machine is developed to infer prior images. The model training is transformed into a bilevel fractional optimization problem by incorporating measurement principles into the data-driven model, automating both model parameter selection and training process. The new algorithm is extensively compared, both qualitatively and quantitatively, with leading reconstruction algorithms. The comparison results show its significant advantages in terms of reconstruction quality and robustness across various imaging conditions. Our research introduces innovative concepts and fresh perspectives for artifact reduction and enhanced noise robustness, potentially improving the effectiveness and applicability of the measurement technique.</p>