Abstract: Objective  To address the issue of noise in medical images, this paper proposes a left ventricular image segmentation method integrated with a denoising module to improve segmentation accuracy. Methods   The denoising module is based on a denoising diffusion probabilistic model, and the segmentation model includes two branches: motion estimation and segmentation. This paper modifies the prediction target of the denoising module to the original signal instead of noise, enabling the end-to-end cascade training process of the denoising module and the segmentation model. Results   On the EchoNet-Dynamic dataset, the segmentation performance of traditional denoising methods was inferior to the benchmark model; the Noise2Noise model showed improvement in three metrics, while our proposed method achieved improvement in all four metrics. On the ACDC dataset, our method outperformed the benchmark model, while other methods either performed worse than the benchmark or showed no statistical difference. Conclusion   Traditional denoising methods can impair segmentation performance, whereas our proposed method can stably and effectively improve segmentation performance. Experiments verify the feasibility and potential clinical application value of the proposed method.

Key words: left ventricular segmentation, denoising model, segmentation model, medical image segmentation, diffusion probabilistic model, cascade training