Title: IntraStyler: Generating Diverse Styles Within a Domain for Cross-Modality MRI Adaptation

Abstract: Accurately segmenting vestibular schwannomas and cochleae in T2 MRI scans is a critical clinical task, though it demands significant manual annotation. To overcome the scarcity of labeled T2 data, researchers frequently employ domain adaptation (DA) techniques to align labeled contrast-enhanced T1 images with unlabeled T2 datasets. However, current approaches primarily concentrate on aligning differences between domains, often neglecting the substantial variability that exists within the target domain itself. Factors such as varying scanner types, magnetic field strengths, and acquisition protocols can cause images from the same domain to differ significantly. By disregarding this intra-domain diversity, existing methods tend to generate uniform synthetic images, which restricts the robustness and generalizability of subsequent segmentation models. To mitigate this limitation, we introduce IntraStyler, a novel 3D unpaired image translation framework. This method automatically identifies fine-grained style variations within the target domain without relying on pre-defined sub-groups, creating diverse synthetic images by applying style references specific to each individual scan. Central to this approach is a 3D style encoder, which utilizes a new contrastive learning objective to isolate style-specific embeddings from anatomical features. Building upon the winning solution of the CrossMoDA challenge, IntraStyler enhances previous efforts by producing a wider variety of synthetic data, thereby improving the reliability of downstream segmentation tasks. The source code is publicly accessible at https://github.com/MedICL-VU/IntraStyler.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC