Hairong Zheng

CSM$_{\text{CCTA}}$ outperformed existing metrics, revealing that specialized models can significantly enhance the accuracy of automated CCTA report generation.

MCSeg outperforms existing methods in cardiac image segmentation by effectively bridging transformer and CNN architectures, showcasing its prowess in few-shot learning scenarios.

VLMs can slash 88% of their visual tokens without performance loss by dropping tokens based on reconstruction error, sidestepping attention-based methods.