Investigating the Agreement with Human Readers and Generalisation Capabilities of a Transfer Learning Approach for Predicting the Malignancy of Solitary Pulmonary Nodules in CT Screening