Effective thermal conductivity is an important property of composites for different thermal management applications. Although physics-based methods, such as effective medium theory and solving partial differential equations, are widely applied to extract effective thermal conductivity, recently there is increasing interest to establish the structure-property linkage through machine learning methods. The prediction accuracy of conventional machine learning methods highly depends on the features (descriptors) selected to represent the microstructures. In comparison, 3D convolutional neural networks (CNNs) can directly extract geometric features of composites, which have been demonstrated to establish structure-property linkages with high accuracy. However, to obtain the 3D microstructure in the composite is challenging in reality. In this work, we use 2D cross-section images and 2D CNNs to predict effective thermal conductivity of 3D composites, since 2D pictures can be much easier to obtain in real applications. The results show that by using multiple cross-section images along or perpendicular to the preferred directionality of the fillers, 2D CNNs can provide quite accurate prediction. Such a result is demonstrated with isotropic particle filled composites and anisotropic stochastic complex composites. In addition, we also discuss how to select representative cross-section images. It is found that the average over multiple images and the use of large-size images can reduce the uncertainty and increase the prediction accuracy. Besides, since cross-section images along the heat flow direction can distinguish between serial structures and parallel structures, they are more representative than cross-section images perpendicular to the heat flow direction.