LowBridge transfers MRI source-modality knowledge to CT via low-level edge features, outperforming ten existing methods in a source-only domain generalization setting
Synopsis
For cross-modal medical image segmentation with MRI-CT transfer, this work proposes LowBridge under a source-only domain generalization setting where training sees only source-modality samples and testing uses unlabeled target-modality images: a generative model is first trained to recover source images from low-level features such as edges, a segmentation model is then trained separately on the generated source images, and at test time edge features from target images are fed to the pretrained generative model to produce source-style target-domain images for segmentation, achieving performance better than ten existing approaches on multiple public datasets, with ablations indicating compatibility with different generative and segmentation models.
Figure 1: Framework of our LowBridge. Low-level features ( e.g. edge features) are treated as domain-invariant representations to train a generative model, 𝒢 \mathcal{G} , to generate source-style images. Edge features extracted from unlabeled target images are then input to 𝒢 \mathcal{G} , followed by a segmentation model, 𝒮 \mathcal{S} , pretrained on the source data, to output the predictions.
arXivInterpretation
It proposes LowBridge, which bridges the inter-domain gap by exploiting the observation that cross-modal images share similar low-level features such as edges, converting target-modality images into a source-modality style before segmentation. Unlike the common practice of training on the source modality and transferring to the target modality, this method uses low-level features as a shared cross-modal interface and relies on a generative model for style conversion, so no target-modality labels are needed during training. The paper reports results on multiple public datasets and states that it outperforms ten existing approaches; the abstract does not give dataset names, sample sizes, or numerical metrics.
The pipeline has two stages: a generative model is trained to recover source images from their edge features, and a segmentation network is then trained separately on the generated source images. Generation and segmentation are decoupled into two independent training stages, so at test time only edge features from target images are needed to generate source-style images for the pretrained segmentation network. The abstract describes the training and testing procedure but does not provide network architectures, loss functions, or training details.
Ablation studies show that LowBridge is compatible with different types of generative and segmentation models. This compatibility suggests the method is not tied to a specific backbone and could benefit from future advances in generative and segmentation models. The abstract states that ablation studies support this conclusion but does not list the model types compared or the specific ablation settings.
Perspective
The work targets a source-only domain generalization setting where only source-modality samples are available during training and unlabeled target-modality images are used at test time, with the abstract explicitly focusing on MRI-CT transfer. It suits research and engineering scenarios that need segmentation without target-modality labels, and because it claims compatibility with different generative and segmentation models, it may benefit as those models advance.
The abstract does not specify the public datasets used, sample sizes, or evaluation metric values, nor does it list the ten compared methods; the types of generative and segmentation models covered by the ablations are also not detailed. Whether the shared low-level feature assumption holds for other modality pairs or anatomical structures still needs more evidence.