Denoising diffusion implicit models and classifier guidance for anomaly detection in MRI images

Abstract

Recently, deep neural networks have demonstrated impressive image synthesis and reconstruction performances, and have been extensively studied in various computer vision tasks. In the medical field, weakly supervised lesion localization is taking a lot of interest as its simple requirement on image- level labeling instead of detailed segmentation which is time-consuming and variable based on different physician annotations. While denoising diffusion probabilistic models have achieved high-quality image generation over variational autoencoders (VAEs) and better diversity of samples compared to generative adversarial networks (GANs), it consumes a lot of sampling time and computational resources. All of these models also have the probability of preserving fine details in the image. To address this problem, we implement Denoising Diffusion Implicit Models for abnormal-to-normal medical image translation and estimate the anomaly map as the subtraction between input and reconstructed image. We use the additional classifier model to effectively localize the tumor region in the image and direct DDIM during the forward sampling process. We evaluate our model combination on BraTS 2021 dataset for lesion localization and compare its performance with the original DDPM, and Fix-Point GAN model. The experimental results show that our method can export high-quality image reconstruction and clarify the importance of the classifier model on the DDIM sampling method.