<p>The diffusion-based conditional denoising paradigm has demonstrated strong performance in medical segmentation tasks. However, due to the inherent uncertainty in the denoising process, existing methods need to aggregate the results of multiple multi-step denoising processes to achieve optimal performance. Recently, flow matching (FM) has emerged as a faster and more effective alternative to diffusion. Yet, simply replacing the diffusion with FM within the same conditional denoising paradigm yields limited gains. In response, we propose a direct mapping strategy that leverages the bijective properties of FM to eliminate the time-consuming and uncertainty problems of the conditional denoising paradigm. The direct mapping model combines segmentation and generation as conjugation tasks. Therefore, we introduce mask augmentation and frequency domain enhancement to indirectly boost segmentation from the generative perspective. Additionally, we propose a self-tuning mechanism that utilizes the generative mode to produce augmented samples, which further improves segmentation accuracy. Our rigorously designed model, MedSegFM, goes beyond the conditional denoising paradigm, unlocking the potential of FM for medical lesion segmentation. MedSegFM achieves state-of-the-art performance on five lesion segmentation tasks across diverse modalities. Notably, the augmented samples can serve as a data augmentation strategy to enhance other segmentation models.</p>