DFTA : Dual‐Flow Trajectory Alignment for Medical Pathology Image Synthesis

Generating medical images with complex anatomical structures frequently faces challenges of morphological distortion and limited diversity. We investigate these issues within an established latent generative pipeline by introducing Dual‐Flow Trajectory Alignment (DFTA), a Flow‐Matching formulation with parallel mask‐flow and image‐flow conditioning. Rather than replacing standard latent encoders, UNet backbones, or ControlNet‐style conditioning, DFTA focuses on how the paired conditional trajectories are regularized and sampled. During training, trajectory consistency guidance uses image flow as a geometric reference for the mask flow; an Online Straight Euler Augmentation (OSEA) mechanism generates intermediate pseudo‐samples. At inference, conditional stochastic flow sampling introduces controlled variation into otherwise deterministic integration. Images generated by our model improve downstream task performance, achieving the best generation metrics and competitive results across downstream classification models. In particular, for medical image segmentation tasks, our method improves the mean Dice coefficient (mDice) and mean Intersection over Union (mIoU) of SegFormer by 1.79% and 1.92%, respectively, and increases the mDice and mIoU of UNet by 2.06% and 2.20%, respectively. The code is available at: https://github.com/yanguanyue/DFTA‐ .