Transfer learning, as a machine learning technique, is based on transferring previously acquired knowledge from one classification task to another. The application of transfer learning in the creation of deep neural networks allows for the newly created model to start with predefined weights instead of randomly assigned weight values, and thus achieved better classification achievements have reduced training time and even in some situations required less training data. In this concrete study focused on a multi-label classification problem of uncovering one or more respiratory diseases or radiological findings in chest X-ray images, a comparison was made between models trained with and without the utilization of transfer learning. Both approaches were supported by a result postprocessing system which was made to rectify possible mistakes which can arise as a byproduct of the nature of multi-label classification. Comparison was made between these models using two modified classification accuracy metrics. Neural networks trained with randomized starting weight values achieved 61% and 94% patient-based and disease-based accuracy, respectively. In contrast, the model which utilized transfer learning by using weights from a network which solved a similar respiratory disease classification problem achieved 64% and 96% patient-based and disease-based classification accuracy, respectively. While the increase of 3% over the stricter accuracy metric does not seem like a large improvement, the amount of mistakes made on a disease to disease basis was reduced by one third, since the number of incorrectly classified instances was reduced from 6% down to 4%. In addition, the models trained from transferred weights required significantly less training time to reach results that are up to standard, requiring 5 h for model training, in contrast to 16 h required for the other approach.