TuLaBM: Tumor-Biased Latent Bridge Matching for Contrast-Enhanced MRI Synthesis

A research team from institutions including the Cleveland Clinic has published TuLaBM (Tumor-Biased Latent Bridge Matching), a novel AI framework for medical image synthesis. The system addresses a critical bottleneck in oncology: the need for gadolinium-based contrast agents (GBCAs) to create contrast-enhanced MRI (CE-MRI) scans for tumor assessment. These agents increase costs and carry safety risks like nephrogenic systemic fibrosis. TuLaBM instead generates synthetic CE-MRI images directly from standard, non-contrast MRI (NC-MRI) scans, offering a safer and more accessible alternative.

Technically, TuLaBM formulates the image translation as a Brownian bridge transport problem within a learned latent space, a method that provides greater stability than older Generative Adversarial Networks (GANs) and is far more efficient than current diffusion models. Its key innovation is the Tumor-Biased Attention Mechanism (TuBAM), which actively amplifies tumor-relevant features during image generation, and a boundary-aware loss function that sharpens tumor margins. This focus ensures clinical fidelity where it matters most.

In validation on the BraTS2023-GLI brain tumor dataset and an in-house liver MRI dataset, TuLaBM consistently outperformed state-of-the-art models on both whole-image and tumor-specific metrics. Crucially, it demonstrated strong generalization, performing well on unseen liver MRI data in both zero-shot and fine-tuned settings. The most practical breakthrough is its speed: TuLaBM achieves inference times under 0.097 seconds per image, making near-real-time synthesis feasible during patient scans.