Dayang Wang

Dayang Wang, Srivathsa Pasumarthi, Greg Zaharchuk et al.

Deep learning (DL) based contrast dose reduction and elimination in MRI imaging is gaining traction, given the detrimental effects of Gadolinium-based Contrast Agents (GBCAs). These DL algorithms are however limited by the availability of high quality low dose datasets. Additionally, different types of GBCAs and pathologies require different dose levels for the DL algorithms to work reliably. In this work, we formulate a novel transformer (Gformer) based iterative modelling approach for the synthesis of images with arbitrary contrast enhancement that corresponds to different dose levels. The proposed Gformer incorporates a sub-sampling based attention mechanism and a rotational shift module that captures the various contrast related features. Quantitative evaluation indicates that the proposed model performs better than other state-of-the-art methods. We further perform quantitative evaluation on downstream tasks such as dose reduction and tumor segmentation to demonstrate the clinical utility.

Dayang Wang, Yongshun Xu, Shuo Han et al.

Low-dose computed tomography (LDCT) reduces the X-ray radiation but compromises image quality with more noises and artifacts. A plethora of transformer models have been developed recently to improve LDCT image quality. However, the success of a transformer model relies on a large amount of paired noisy and clean data, which is often unavailable in clinical applications. In computer vision and natural language processing fields, masked autoencoders (MAE) have been proposed as an effective label-free self-pretraining method for transformers, due to its excellent feature representation ability. Here, we redesign the classical encoder-decoder learning model to match the denoising task and apply it to LDCT denoising problem. The MAE can leverage the unlabeled data and facilitate structural preservation for the LDCT denoising model when ground truth data are missing. Experiments on the Mayo dataset validate that the MAE can boost the transformer's denoising performance and relieve the dependence on the ground truth data.

Fenglei Fan, Juntong Fan, Dayang Wang et al.

The rapid growth of large models' size has far outpaced that of computing resources. To bridge this gap, encouraged by the parsimonious relationship between genotype and phenotype in the brain's growth and development, we propose the so-called Hyper-Compression that turns the model compression into the issue of parameter representation via a hyperfunction. Specifically, it is known that the trajectory of some low-dimensional dynamic systems can fill the high-dimensional space eventually. Thus, Hyper-Compression, using these dynamic systems as the hyperfunctions, represents the parameters of the target network by their corresponding composition number or trajectory length. This suggests a novel mechanism for model compression, substantially different from the existing pruning, quantization, distillation, and decomposition. Along this direction, we methodologically identify a suitable dynamic system with the irrational winding as the hyperfunction and theoretically derive its associated error bound. Next, guided by our theoretical insights, we propose several engineering twists to make the Hyper-Compression pragmatic and effective. Lastly, systematic and comprehensive experiments on \textcolor{black}{NLP models such as LLaMA and Qwen series and vision models} confirm that Hyper-Compression enjoys the following \textbf{PNAS} merits: 1) \textbf{P}referable compression ratio; 2) \textbf{N}o post-hoc retraining; 3) \textbf{A}ffordable inference time; and 4) \textbf{S}hort compression time. It compresses LLaMA2-7B in an hour and achieves close-to-int4-quantization performance, without retraining and with a performance drop of less than 1\%. We have open-sourced our code in https://github.com/Juntongkuki/Hyper-Compression.git for free download and evaluation.