Damudi Mohammed Z, Kini Anita S
Department of Computer Science & Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education (MAHE) 576104, Manipal, Karnataka, India.
Int J Biomed Imaging. 2024 Dec 19;2024:2352602. doi: 10.1155/ijbi/2352602. eCollection 2024.
Generative models, especially diffusion models, have gained traction in image generation for their high-quality image synthesis, surpassing generative adversarial networks (GANs). They have shown to excel in anomaly detection by modeling healthy reference data for scoring anomalies. However, one major disadvantage of these models is its sampling speed, which so far has made it unsuitable for use in time-sensitive scenarios. The time taken to generate a single image using the iterative sampling procedure introduced in denoising diffusion probabilistic model (DDPM) is quite significant. To address this, we propose a novel single-step sampling procedure that hugely improves the sampling speed while generating images of comparable quality. While DDPMs usually denoise images containing pure noise to generate an original image, we utilize a partial diffusion approach to preserve the image structure. In anomaly detection, we want the reconstructed image to have a structure similar to the original anomalous image, so that we can compare the pixel-level difference between them in order to segment the anomaly. The original DDPM algorithm suggests an iterative sampling procedure where the model slowly reduces the noise, until we have a noise-free image. Our single-step sampling approach attempts to remove all the noise in the image within a single step, while still being able to repair the anomaly and achieve comparable results. The output is a binary image showing the predicted anomalous regions, which is then compared to the ground truth to evaluate its segmentation performance. We find that, while it does achieve slightly better anomaly masks, the main improvement is in sampling speed, where our approach was found to perform significantly faster as compared to the iterative procedure. Our work is mainly focused on anomaly detection in brain MR volumes, and therefore, this approach could be used by radiologists in a clinical setting to find anomalies in large quantities of brain MRI.
生成模型，尤其是扩散模型，因其高质量的图像合成在图像生成领域受到关注，超越了生成对抗网络（GAN）。它们通过对健康参考数据进行建模以对异常进行评分，在异常检测方面表现出色。然而，这些模型的一个主要缺点是其采样速度，到目前为止，这使得它们不适用于对时间敏感的场景。使用去噪扩散概率模型（DDPM）中引入的迭代采样过程生成单个图像所花费的时间相当长。为了解决这个问题，我们提出了一种新颖的单步采样过程，该过程在生成质量相当的图像时极大地提高了采样速度。虽然DDPM通常对包含纯噪声的图像进行去噪以生成原始图像，但我们采用部分扩散方法来保留图像结构。在异常检测中，我们希望重建的图像具有与原始异常图像相似的结构，以便我们可以比较它们之间的像素级差异以分割异常。原始的DDPM算法建议采用迭代采样过程，即模型缓慢降低噪声，直到我们得到无噪声的图像。我们的单步采样方法试图在一步内去除图像中的所有噪声，同时仍然能够修复异常并取得相当的结果。输出是一个显示预测异常区域的二值图像，然后将其与地面真值进行比较以评估其分割性能。我们发现，虽然它确实能实现稍好的异常掩码，但主要改进在于采样速度，我们的方法与迭代过程相比，执行速度明显更快。我们的工作主要集中在脑部磁共振体积的异常检测上，因此，这种方法可被放射科医生在临床环境中用于在大量脑部MRI中发现异常。
