Crack analysis is a crucial component of structural health monitoring. A fundamental aspect of this is the identification and quantification of individual crack instances through crack instance segmentation. Due to the visually linear/curvilinear structure of the cracks, data imbalance problems inevitably arise in the instance segmentation task. To address this, a Crack Causal Augmentation Framework (CCAF) which incorporates morphological dilation and erosion is designed to enhance the representation of cracks within the training images. Furthermore, it has been observed that crack data is highly susceptible to variations in threshold during binarization, to address this issue, Dynamic Binary Threshold (DBT) branch has been implemented to enhance the segmentation accuracy by performing a regression task using spatial features in the segmentation network to determine the optimal binary threshold at the image level. Experimental results demonstrate that the implementation of the CCAF in conjunction with the DBT branch results in a synergistic enhancement of performance. We establish the efficacy of our proposed methods through evaluations on four crack datasets with varying sources and distributions, including a dataset with an extremely low density of crack pixels, where other prevalent crack instance segmentation techniques have been inadequate.