Current mainstream of CT reconstruction methods based on deep learning usually needs to fix the scanning geometry and dose level, which will significantly aggravate the training cost and need more training data for clinical application. In this paper, we propose a parameter-dependent framework (PDF) which trains data with multiple scanning geometries and dose levels simultaneously. In the proposed PDF, the geometry and dose level are parameterized and fed into two multi-layer perceptrons (MLPs). The MLPs are leveraged to modulate the feature maps of CT reconstruction network, which condition the network outputs on different scanning geometries and dose levels. The experiments show that our proposed method can obtain competing performance similar to the original network trained with specific geometry and dose level, which can efficiently save the extra training cost for multiple scanning geometries and dose levels.