Echo from Noise: Synthetic Ultrasound Image Generation Using Diffusion Models for Real Image Segmentation

Abstract

We propose a novel pipeline for the generation of synthetic ultrasound images via Denoising Diffusion Probabilistic Models (DDPMs) guided by cardiac semantic label maps. We show that these synthetic images can serve as a viable substitute for real data in the training of deep-learning models for ultrasound image analysis tasks such as cardiac segmentation. To demonstrate the effectiveness of this approach, we generated synthetic 2D echocardiograms and trained a neural network for segmenting the left ventricle and left atrium. The performance of the network trained on exclusively synthetic images was evaluated on an unseen dataset of real images and yielded mean Dice scores of $$88.6 \pm 4.91$$ 88.6 ± 4.91 , $$91.9 \pm 4.22$$ 91.9 ± 4.22 , $$85.2 \pm 4.83$$ 85.2 ± 4.83 % for left ventricular endocardium, epicardium and left atrial segmentation respectively. This represents a relative increase of 9.2, 3.3 and 13.9% in Dice scores compared to the previous state-of-the-art. The proposed pipeline has potential for application to a wide range of other tasks across various medical imaging modalities.

Synthetic Echocardiograms Generation Using Diffusion Models

“…We first compare our method with the Echo from Noise method proposed by Stojanovski et al [18]. This method is limited to the generation of one 2D echocardiogram image, not a sequence, whereas we generate echocardiogram sequences with cardiac motion [18]. We only use the FID score for comparison here because Echo from Noise only creates a single frame.…”

“…We only use the FID score for comparison here because Echo from Noise only creates a single frame. We used the real and synthetic images provided by [18] to estimate the FID score of the Echo from Noise method. The FID score of echocardiogram images generated using Echo from Noise is 75.36 while FID score of images generated using our method is 17.65.…”

“…Diffusion models have been used to generate images [18], [24], [25], and speech signals [26]. Recently, several methods have been proposed which explore the generation of video sequences [27]- [29].…”

“…Diffusion models have been proposed for generating echocardiograms [18], [31]. Stojanovski et al [18] proposed "Echo from Noise" to produce one annotated image of an echocardiogram. This method does not produce echocardiogram sequences.…”

“…One way to address the problems above is to use synthetic data for echocardiogram analysis. Generative machine learning methods, such as GANs [16] or diffusion models [17] have been used to generate synthetic data that can be used for training [18]- [22]. In this paper, we propose a video diffusion method to generate echocardiograms.…”