Co-occurring diseases heavily influence the performance of weakly supervised learning models for classification of chest CT