Medical image semantic segmentation commonly uses fully-supervised learning. However, its requirement to use all labeled training images requires a lot of resources and costs. Semi-supervised learning is proposed to tackle this problem. But, medical image segmentation is frequently faced with a few amounts of training images, especially in the specific modality. This research focuses on implementing the cross-modality concept in semi-supervised image segmentation. The method generally consists of data augmentation and two phases of learning. Data augmentation uses task-driven and semi-supervised techniques. Cross-modality is implemented in the third phase of learning to synthesize the image from assistant images. Hence, the cross-modality concept makes the assistant modality images leverage the training phases. The system is evaluated using the Dice Score and Volumetric similarity. The experiment result shows that the cross-modality concept’s integration enhances the semi-supervised image segmentation task. The enhancement also causes a reduction in accuracy degradation.