Recurrent Visual Feature Extraction and Stereo Attentions for CT Report Generation