Rare existing studies in Instance Segmentation (IS) paid attention to the issue of image scarcity in small-sample domains. In particular, we observe three challenges presented in publicly available IS data sets in these domains. First, the background in these data sets varies greatly and unseen contexts may cause significant distribution shifts. Second, mis-segmented regions are usually correlated with target regions highly. Third, objects contained in different categories often share visual similarities, leading to misclassifications. Although Copy-Paste and InstaBoost have been proposed for IS tasks, they did not take these challenges into consideration. To address those challenges, we introduce a data augmentation approach, namely, Instance Segmentation Adapted to Small-Sample Domains, or IS-SSD for short, to boost the performance of IS methods. The IS-SSD comprises a Synthetic Image Generator (SIG) and a Target Part Classifier (TPC). The SIG synthesizes new training images by compositing the foreground regions that are prone to be incorrectly segmented from an image to a different image. As a result, the model trained using these images will learn complex foreground-background relationships. On the other hand, the TPC aims to mitigate spurious activations within object regions, to improve the integrity of the object detected. Experimental results show that our method has achieved greater performance gains in small-sample instance segmentation, compared to baselines. We believe that the promising results are due to the ability of the IS-SSD to increase the diversity of training images and improve the target activation accuracy.