1. How is stroke lesion volume measured?
Stroke lesion volume is typically measured on post-treatment Non-Contrast CT (NCCT) scans. Manual contouring of stroke lesions is considered the gold standard for volume measurement, but it is time-consuming and observer-dependent. Despite attempts to automate segmentation, there are no well-established methods for NCCT due to limitations such as low signal-to-noise ratio, low contrast of soft tissues, partial volume effects, and scanner variability. The proposed Diffusion Probabilistic Model (DPM) based approach aims to accurately segment ischemic or hemorrhagic lesions on NCCT scans.
read more
2. What is the diffusion model in image generation?
The diffusion model (DM) in image generation is a LVM that has been successfully applied to the field of image generation. It involves a reverse process modeled by a first-order Markov chain, starting at x T ~ N (0, I), and a diffusion process q(x 1:T |x 0) that adds standard normal noise with a variance schedule b t. The joint distribution p(x 0:T) is called the reverse process, and the diffusion process q(x 1:T |x 0) is also a Markov chain. The goal is to minimize the Kullback-Leibler (KL) Divergence between q(x 1:T |x 0) and p(x 0:T) to optimize the variational boundary on the negative log-likelihood. The optimization process involves predicting the noise e t from the diffused image x t using a neural network. The diffusion process is an approximate conditional posterior q(y 1:T |y 0, x 0:T), and the loss function is defined as minimizing the KL divergence between the conditional posterior and the unnormalized distribution. The model can infer back and get the label y 0 by estimating the noisy label y T from the image x T. However, it is challenging to predict y T accurately due to the severe destruction of image information. To obtain a good initial y t for inferring y 0, a time window of length T p is introduced, and the loss function is further simplified. The model assumes that images and labels are overlain by the same Gaussian noise during the diffusion process, allowing for the sampling of images and labels at time t. The loss function L d0 optimizes the last step of generating y 0, and a self-attention module with a convolutional layer of frozen parameters is used to remove cumbersome discrete segmentation points.
read more
3. What is the transition term in SDPM?
The transition term in SDPM is a Gaussian distribution, N u th (y t, t), with trainable parameters. It is suggested in [13] that s 2 t can be either b t or bt. This term plays a crucial role in the reverse process of SDPM, allowing the model to predict mt given the label y 0 with the shared e t in the posterior. The final label y 0 can then be inferred through a Markov process. To reduce uncertainty in inferring y 0, SDPM adds another loss function to restrict the difference between the true and predicted final labels y 0. This loss function, L d, is reparameterized and optimized along with L d1 and L d2. The model also utilizes x t to predict the shared noise et (x t, t) by leveraging the network predicting the noise at each diffusion process, which carries the label information. Overall, the transition term and loss functions in SDPM contribute to improving the segmentation performance.
read more
4. How can the final label y0 be inferred using SDPM?
The final label y0 can be inferred in four ways. The first method involves directly outputting yavg 0 = y0, where y0 is the output of the trained network given the clean image x0. The second method utilizes the salient weight ps t over the time window Ti, represented as ysal 0 = 1 N N n=1 Ti<T t=0 y0 ps t, where ps t = 1 - (t Ti) n, with n > 1 being gradually degraded coefficients. The third method employs Markov chain inference starting at time Ti, denoted as yinfer 0 = 1 N N n=1 IL(d i , Ti), where d i is in the range (0,1) and Ti is within the time range {0, *, *, *}. These methods provide different approaches to infer the final label y0 in the SDPM process.
read more