Unsupervised deep learning enables low-dose extreme sparse view HAADF-STEM-EDX tomography reconstruction.
AI model for accurate classification of the mucosal lesion along with depth preidction using a single endoscopic image was proposed. Results show that the developed model is on par with or even exceeds performance of practiced experts.
Thorough review on unsupervised deep learning methods including self-supervised methods and generative model-based methods for biological image reconstruction and enhancement is provided.
Resolution of STEM-EDX tomography is greatly enhanced through unsupervised deep learning. We verified the method by comparing the optical properties.
A novel unpaired learning scheme for accelerated MRI, OT-cycleGAN was extensively applied and was found effective for the reconstruction of multi-coil static MRI.
OT-cycleGAN for the reconstruction of time resolved magnetic resonance angiography (MRA) was proposed. The derived method enables flexible control of sptial and temporal resolution.
SliceGAN, stochastic model to synthesize 3D microstructures from 2D images, is endowed with the ability to disentangle features, and continuously control these features via AdaIN.
Two-stage unsupervised reconstruction method for 3D TOF-MRA is developed. A novel projection discriminator in the axial reconstruction step drastically enhances the vessel visiblity.
BarbellNet, which consists of long stack of residual channel attention block(RCAB) was proposed for the reconstruction of fast MRI reconstruction. Reconstruction results through this model was placed 6th in the NeurIPS2020 fastMRI challenge.
Unsupervised missing-cone resolving method is proposed in the context of ODT.