Weakly-Supervised Multimodal Learning on MIMIC-CXR