Recently, deep learning-enabled joint-source channel coding (JSCC) has received increasing attention due to its great success in image transmission. However, most existing JSCC studies only focus on single-input single-output (SISO) channels. In this paper, we first propose a JSCC system for wireless image transmission over multiple-input multiple-output (MIMO) channels. As the complexity of an image determines its reconstruction difficulty, the JSCC achieves quite different reconstruction performances on different images. Moreover, we observe that the images with higher reconstruction qualities are generally more robust to the noise, and can be allocated with less communication resources than the images with lower reconstruction qualities. Based on this observation, we propose an adaptive channel state information (CSI) feedback scheme for precoding, which improves the effectiveness by adjusting the feedback overhead. In particular, we develop a performance evaluator to predict the reconstruction quality of each image, so that the proposed scheme can adaptively decrease the CSI feedback overhead for the transmitted images with high predicted reconstruction qualities in the JSCC system. We perform experiments to demonstrate that the proposed scheme can significantly improve the image transmission performance with much-reduced feedback overhead.