Adaptive Convolutional Dictionary Network for CT Metal Artifact Reduction

Abstract

Action segmentation, inferring temporal positions of human actions in an untrimmed video, is an important prerequisite for various video understanding tasks. Recently, unsupervised action segmentation (UAS) has emerged as a more challenging task due to the unavailability of frame-level annotations. Existing clustering- or prediction-based UAS approaches suffer from either over-segmentation or overfitting, leading to unsatisfactory results. To address those problems,we propose Predictive And Contrastive Embedding (PACE), a unified UAS framework leveraging both predictability and similarity information for more accurate action segmentation. On the basis of an auto-regressive transformer encoder, predictive embeddings are learned by exploiting the predictability of video context, while contrastive embeddings are generated by leveraging the similarity of adjacent short video clips. Extensive experiments on three challenging benchmarks demonstrate the superiority of our method, with up to 26.9% improvements in F1-score over the state of the art.

b-MAR: bidirectional artifact representations learning framework for metal artifact reduction in dental CBCT

“…CNNMAR employs a convolutional neural network to generate a prior image in the image domain, aiding in sinogram domain correction (Zhang and Yu 2018). ACDNet encodes the prior structure of metal artifacts into an explicit weighted convolutional dictionary, guiding the network in artifact removal (Wang et al 2022). Regarding unsupervised deep MAR methods, we select the classic ADN algorithm, which disentangles artifacts and texture content in the latent space of CT images (Liao et al 2019).…”

“…For example, Park et al repaired inconsistent sinogram by removing the primary metalinduced beam-hardening factors along the metal trace in the sinogram (Park et al 2018). Wang et al designed an image-domain adaptive convolutional dictionary network which used prior information that metal artifacts present the form of non-local repeating stripes (Wang et al 2022). Lin et al construct two cascaded sub-networks in the image domain and sinogram domain respectively.…”

Mud-Net: multi-domain deep unrolling network for simultaneous sparse-view and metal artifact reduction in computed tomography

“…The corresponding artifact-affected CT images are reconstructed from the sparse-view sinograms by FBP, and all CT images are resized to 416 × 416. In particular, for the full-view MAR task, we use the sinogram size consistent with existing MAR methods [16,30]. In order to validate the generalization ability of the method, our method was qualitatively compared with the existing MAR methods under the full-view condition on clinical data.…”

“…To demonstrate the superiority of our Mud-Net in full-view CT, we compared Mud-Net with existing MAR methods, including LI [18], NMAR [19], CNNMAR [23], DuDoNet [26], DSCMAR [27], InDuDoNet [30], InDuDoNet+ [33], DICDNet [31], and ACDNet [16]. Table 1 summarizes the average PSNR and SSIM values on varying metal sizes under the full-view CT system.…”

“…In fact, the effectiveness and superiority of dual-domain knowledge have been verified in various tasks of medical image processing [13][14][15][16]. It is clear that making the most out of the structural information of domain knowledge greatly improves the correction accuracy.…”