Cross-Modality Medical Image Segmentation

The field develops methods to segment organs, lesions, and other structures when images come from different modalities, such as MRI, CT, PET, and ultrasound, or when some modalities are missing. Common approaches include unsupervised domain adaptation, image synthesis, multimodal feature fusion, self-supervised pretraining, and anatomical shape or topology priors. Models are tested on settings such as brain-tumor MRI, cardiac MRI-to-CT transfer, and PET/CT lesion segmentation, often with scarce labels or shifts between hospitals. A central challenge is that complementary modalities can improve predictions, but misalignment, noise, or conflicting signals can make fusion worse than a strong single-modality model, so external validation and modality-specific comparisons remain important.