USB: Unified Synthetic Brain Framework for Bidirectional Pathology-Healthy Generation and Editing

TL;DR

USB addresses the challenge of scarce paired pathological–healthy brain data by introducing a unified bidirectional framework that jointly models lesions and brain anatomy via a paired diffusion process. It presents Anatomy Consistency Guidance and Lesion Consistency Guidance to preserve global brain structure while ensuring lesion localization during editing, enabling both healthy-to-pathology synthesis and pathology-to-healthy restoration within a single architecture. The approach is validated across six public brain MRI datasets, achieving state-of-the-art performance on generation and editing tasks and establishing the first unified benchmark for bidirectional brain image generation-editing. This work enables scalable dataset creation and robust neuroimaging analysis, with code released for broader adoption in clinical research and potential cross-modality extensions.

Abstract

Understanding the relationship between pathological and healthy brain structures is fundamental to neuroimaging, connecting disease diagnosis and detection with modeling, prediction, and treatment planning. However, paired pathological-healthy data are extremely difficult to obtain, as they rely on pre- and post-treatment imaging, constrained by clinical outcomes and longitudinal data availability. Consequently, most existing brain image generation and editing methods focus on visual quality yet remain domain-specific, treating pathological and healthy image modeling independently. We introduce USB (Unified Synthetic Brain), the first end-to-end framework that unifies bidirectional generation and editing of pathological and healthy brain images. USB models the joint distribution of lesions and brain anatomy through a paired diffusion mechanism and achieves both pathological and healthy image generation. A consistency guidance algorithm further preserves anatomical consistency and lesion correspondence during bidirectional pathology-healthy editing. Extensive experiments on six public brain MRI datasets including healthy controls, stroke, and Alzheimer's patients, demonstrate USB's ability to produce diverse and realistic results. By establishing the first unified benchmark for brain image generation and editing, USB opens opportunities for scalable dataset creation and robust neuroimaging analysis. Code is available at https://github.com/jhuldr/USB.

Figures (17)

• Figure 1: Four unified tasks supported by USB.(a) Unconditional generation of paired brain image and lesion mask; (b) Conditional generation given a specific lesion mask; (c) Pathology-to-healthy editing that reconstructs a lesion-containing brain into a healthy brain; (d) Healthy-to-pathology editing that embeds a given lesion mask into a healthy brain.
• Figure 2: Framework of USB, illustrating the overall architecture of our unified synthetic brain framework.(top-left) The bidirectional training process of USB. (top-right) Unconditional generation of paired lesion–brain image. (bottom-left) Conditional brain image generation given a lesion mask. (bottom-middle) Pathology-to-healthy editing given a pathology brain image with lesion embedded. (bottom-right) Healthy-to-pathology editing given a healthy brain image and a lesion mask.
• Figure 3: Consistency Guidance Algorithm during editing.(top) Lesion Consistency Guidance applied in the healthy-to-pathology editing process. (bottom) Anatomy Consistency Guidance applied during both pathology-healthy and healthy-to-pathology editing.
• Figure 4: Qualitative results of USB on generation tasks.(a) unconditional generation, illustrating USB's capability to generate lesion masks with diverse locations and complex shapes, together with their corresponding pathological brain images; (b) conditional generation, showing USB's ability to generate unlimited pathological brain images conditioned on varying lesion masks.
• Figure 5: Comparison of bidirectional brain editing.(a) pathology-to-healthy, the circles and arrows highlight lesion regions and unsuccessful reconstructions; (b) healthy-to-pathology. Note that SynthSR and Brain-IDcannot perform healthy-to-pathology editing.