Figure 1:

Photographic images typically have objects with discriminative parts (e.g., wheels of a bicycle) at the center, enabling instance discrimination methods to learn generalizable representations; medical images generated from a particular imaging protocol exhibit remarkable resemblance in anatomy (e.g., lungs), hindering these methods from capturing distinct features, resulting in poor transferability. To address this issue, this paper empowers various instance discrimination methods with the fine-grained, local context information embedded in medical images.