Introduction
Generative AI, particularly diffusion models, has demonstrated significant potential for synthesizing high-fidelity 3D medical images. Such synthetic data is vital for training robust AI models in data-scarce environments, preserving patient privacy, and simulating rare clinical anomalies. However, existing 3D medical diffusion models are often limited by narrow generalizability, prohibitive inference times (often exceeding 10 minutes for a single volume), and weak alignment between generated anatomy and input conditioning masks.
MAISI-v2 addresses these challenges by evolving the original MAISI framework into a faster, more memory-efficient system that maintains strict obedience to input conditions, particularly for small abnormalities like tumors.
Methodology
MAISI-v2 introduces two primary methodological advancements over its predecessor: the adoption of Rectified Flow and a novel contrastive loss function.
Rectified Flow-based Latent Diffusion
The framework replaces the standard stochastic Denoising Diffusion Probabilistic Model (DDPM) with a Rectified Flow model. Unlike the curved, noisy trajectories of traditional diffusion, Rectified Flow learns a deterministic, straight-line Ordinary Differential Equation (ODE) path from noise to image. This transition allows for high-quality image generation in significantly fewer steps, drastically reducing the computational burden of 3D synthesis.
Region-Specific Contrastive Loss
While Rectified Flow increases speed, its deterministic nature can reduce robustness to fine-grained conditioning compared to stochastic DDPM. To ensure the model accurately reflects small segmentation masks (e.g., tumors), the authors implemented a Region-Specific Contrastive Loss.
As shown in the figure above, the model processes two inputs derived from the same noise: one with the target tumor mask and one with a "foreground-free" perturbed mask. The loss function enforces two constraints:
- ROI Sensitivity: The generated outputs must differ significantly within the region of interest (the tumor).
- Background Consistency: The generated outputs must remain identical in all other regions.
Experimental Setup
The model was trained on a massive dataset of over 12,000 CT scans (~107k 3D images) aggregated from various public sources (e.g., AbdomenCT-1K, AMOS22, LIDC) and in-house datasets. Evaluation was conducted using the AutoPET2023 dataset for out-of-distribution (OOD) image quality and several Medical Segmentation Decathlon (MSD) tasks to assess downstream performance.
Results and Discussion
MAISI-v2 achieves a 33x acceleration in generation speed, requiring only 30 steps compared to the 1000 steps of the original DDPM-based MAISI. A high-resolution $512 \times 512 \times 512$ volume can be generated in just 21 seconds on a single H100 GPU.
Qualitative results demonstrate that MAISI-v2 produces sharper, more anatomically consistent images than previous methods like HA-GAN or MedSyn, which often exhibit blurring or structural artifacts. Quantitatively, the model achieved an average FID of 2.322 on the unseen AutoPET2023 dataset. Furthermore, augmenting training sets with MAISI-v2 synthetic data led to significant improvements in 3D segmentation performance, with Dice scores increasing by up to 7.5% for lung tumors and 6.4% for pancreatic tumors compared to training on real data alone.
Conclusion
By combining Rectified Flow with a novel contrastive loss, MAISI-v2 provides a practical solution for high-speed, high-fidelity 3D medical image synthesis. While currently focused on CT imaging, future iterations aim to extend these capabilities to MRI and PET modalities, further broadening its clinical utility for data augmentation and rare case simulation.
Summary updated