Generation of Musculoskeletal Ultrasound Images with Diffusion Models

The recent advances in deep learning have revolutionised computer-aided diagnosis medical imaging. However, approaches to unveil their full potential require significant amounts of data, which can be a challenging task some scientific fields, such as musculoskeletal ultrasound imaging, data privacy and security reasons lead important limitations the acquisition distribution process patients’ data. For this reason, different generative methods been introduced significantly reduce required amount real by generating synthetic images, almost indistinguishable from ones. In study, power diffusion models is incorporated for generation realistic small set images four muscles. Afterwards, similarity generated assessed with types qualitative quantitative metrics that correspond well human judgement. particular, histograms pixel intensities two sets demonstrated distributions are statistically similar. Additionally, well-established LPIPS, SSIM, FID, PSNR used quantify these images. achieved extremely high scores all metrics. Subsequently, high-level features extracted visualized two-dimensional space inspection structure identify patterns. From representation, hard distinguish. Finally, we perform series experiments assess impact training highly efficient Attention-UNet clinical application muscle thickness measurement. Our results depict play role model’s final performance improvement systems ultrasound.