“…Transfer Learning (TF) strategies are commonly adopted when limited data is available, e.g., medical images. Here, CNNs are trained on a large-scale selection of natural images and the learned knowledge is transferred and repurposed for the new task (Lopes & Valiati, 2017;Rajaraman et al, 2018a;Rajaraman et al, 2018b). However, unlike natural images, medical images exhibit different visual characteristics including color, texture, shape, appearance, and their combinations, and exhibit low intra-class variance and high inter-class similarity (Suzuki, 2017), e.g., CXRs.…”