DescriptionIn medical image computing, the problem of heterogeneous domain shift is quite common and severe, causing many deep convolutional networks to under-perform on various imaging modalities. Retraining the network is difficult since annotating the new domain data is prohibitively expensive, specifically in medical areas that require expertise. While recent works show approaches to tackle this problem using unsupervised domain adaptation, segmentation modules in such methods can be improved vastly. Our implementation provides a segmentation improvement on the current state-of-the-art framework, Synergistic Image and Feature Adaptation(SIFA). We revisit atrous spatial pyramid pooling while using convolutional features as well as image features for multi-scale object segmentation. We have validated the effectiveness of the improvement on the framework using the challenging application of cross-modality segmentation of cardiac structures. To demonstrate the robustness of the module, extensive experiments have been performed on Long-Axis(MMWHS) cross-modal cardiac segmentation tasks.