Abstract
As whole-body multimodal medical imaging scanners gain recognition for their effectiveness in medical applications, the excessive acquisition time in PET-MR scanning remains a significant barrier to more efficient clinical practice. Deep learning-based MRI translation offers a potential solution to reduce scan duration. However, current models often focus on specific anatomical regions and face challenges associated with highly heterogeneous feature distributions in whole-body scans, primarily due to varying anatomical regions and the presence of lesions or pathological tissues. This paper addresses these challenges through a novel Heterogeneity-Adaptive Diffusion Schrödinger Bridge (HA-DSB) framework.
HA-DSB explicitly models translation as stochastic transport between source and target distributions, incorporating region context embeddings derived from a vision-language model (VLM) to enable region-specific modeling. To enhance the fidelity of pathological tissues, lesion-aware metabolic priors from PET are directly integrated into the bridge dynamics via a dual-stage guidance mechanism. Specifically, a PET-guided noise modulation module adaptively scales spatial diffusion perturbations during the forward process, while PET features are leveraged in the reverse process to selectively amplify lesion-relevant structures through an attention mechanism.
Experiments demonstrate the superiority of our method across different body regions in whole-body MRI translation, showing improved translation quality in lesion areas under PET guidance. Our code is available on GitHub.
Blogger's Review: This paper successfully tackles the heterogeneity issue in whole-body MRI translation with the HA-DSB framework, showing significant improvements in lesion areas. The innovative integration of PET information highlights the potential and future applications of deep learning in medical imaging, paving the way for enhanced clinical practices.