Jian-Qing Zheng

Jian-Qing Zheng

Postdoc @ Oxford




Selected Research Projects
[]

a. Biomedical Data Analysis

Preprint
sym sym

Deformation-Recovery Diffusion Model (DRDM): Instance Deformation for Image Manipulation and Synthesis

Website arXiv code

J. Q. Zheng†, Y. Mo†, Y. Sun, J. Li, F. Wu, Z. Wang, T. Vincent, B. W. Papiez

  • This is the first study to explore diverse deformation generation for one specific image without an atlas image.
  • - A novel diffusion model is proposed based on deformation diffusion and recovery.
  • - The model generates diverse and realistic deformation for each instance image.
  • - A random and reasonable deformation field sampling method is proposed for training.
  • - The model enables and improves data augmentation in few-shot segmentation.
  • - The model enables and improves data synthesis for registration model training.
Med Image Anal (IF:13.8)
sym sym

Residual Aligner-based Network (RAN): Motion-Separable Structure for Coarse-to-fine Discontinuous Deformable Registration

Website DOI arXiv code
CSDN博客 CVer公众号

J. Q. Zheng, Z. Wang, B. Huang, N. H. Lim, B. W. Papiez

  • This is the first quantitative investigation of deep learning methods for discontinuous deformable registration.
  • - The proposed RAN obtains the state-of-the-art accuracy and better efficiency.
  • - Motion separability quantifies upper limit of the predictable motions' difference.
  • - Motion-Separable structure is used for a higher motion separability.
  • - Multi-head mask is used to disentangle the motions of different organs.
  • - Confidence weight is estimated to the weight and refine the predicted motions.
CMIG (IF:7.4) + MICCAIW 2020
sym sym

Accurate volume alignment of arbitrarily oriented tibiae based on a mutual attention network for osteoarthritis analysis | DOI Patent

D-net: siamese based network for arbitrarily oriented volume alignment | DOI arXiv

J. Q. Zheng, N. H. Lim, B. W. Papiez

  • It is the earliest work that applies the attention mechanism to image registration, enabling sub-voxel errors even from an arbitrary initial orientation.
  • - Alignment of arbitrary oriented CT volumes allows extraction of contrasted cartilage.
  • - Only proposed D-net achieved accurate alignment of arbitrary oriented volumes.
  • - D-net extracts the common features and the spatial information via Siamese structure.
  • - Mutual attention enhances long-range connection between two-branch D-net structure.
  • - The alignment framework may be generalizable for multiple longitudinal studies.
ICRA 2020
sym sym

ACNN: a Full Resolution DCNN for Medical Image Segmentation

DOI arXiv

X. Y. Zhou†, J. Q. Zheng†, P. Li, G. Z. Yang

  • A pooling-free network structure is proposed to achieve full-resolution feature processing using a theoretically optimal dilation setting for a larger receptive field, which achieves a higher accuracy even with fewer parameters.
  • - The left figure shows the receptive field with the dilation setting (1,2,5...) used for DeepLab.
  • - The right figure shows the receptive field with the dilation setting (1,3,9...) used for our ACNN.
  • - Our setting achieves larger receptive field with the same number of layers.

b. Surgical Navigation

ICRA 2019
sym sym

Towards 3D Path Planning from a Single 2D Fluoroscopic Image for Fenestrated Endovascular Aortic Repair

DOI arXiv code

J. Q. Zheng, X. Y. Zhou, C. Riga, G. Z. Yang

  • The first real-time framework of 3D intra-operative AAA skeleton instantiation from a single 2D AAA fluoroscopic image is proposed for 3D robotic path planning, achieving sub-pixel error in the clinical data.
  • - Left: the 3D abdominal aortic center line recovered from one 2D X-ray image.
  • - Right: how it looks like when a catheter moves along the recovered center line of aorta.
MICCAI 2022
sym

Self-Supervised Depth Estimation in Laparoscopic Image using 3D Geometric Consistency

DOI arXiv code code

B. Huang, J. Q. Zheng, A. Nguyen, C. Xu, I. Gkouzionis, K. Vyas, D. Tuch, S. Giannarou, D. S. Elson

  • - Most recent depth estimation work focused on the left-right consistency in 2D and ignored valuable inherent 3D information on the object in real world coordinates, meaning that the left-right 3D geometric structural consistency is not fully utilized.
  • - To overcome this limitation, we present M3Depth, a self-supervised depth estimator to leverage 3D geometric structural information hidden in stereo pairs while keeping monocular inference.
  • - The method also removes the influence of border regions unseen in at least one of the stereo images via masking, to enhance the correspondences between left and right images in overlapping areas.
  • - Extensive experiments show that our method outperforms previous self-supervised approaches on both a public dataset and a newly acquired dataset by a large margin, indicating a good generalization across different samples and laparoscopes.

c. Computational Biology and Bioinformatics

Cell Research (IF:46.3)
sym

Deep learning-based rapid generation of broadly reactive antibodies against SARS-CoV-2 and its Omicron variant

Website DOI GitHub stars Explore XBCR-net in Colab
iNature公众号 简书
eng.ox.ac.uk news.nankai.edu 今日头条 搜狐网

H. Lou†, J. Q. Zheng†, X. Fang, Z. Liang, M. Zhang, Y. Chen, C. Wang, and X. Cao.

  • The first deep learning method is proposed for predicting cross-reactive antibodies, using ACNN structure for feature extraction.
  • a. The features of the amino acid sequences of VH, VL and RBD sequences were extracted, localized and max-pooled to be concatenated together as input to the fully connected layers. The active features in the latent space were then processed by Multi-Layer Perceptron to predict the binding probability of antibody to multiple antigens. The impact score of VH, VL and RBD is calculated on the local histogram impact score map, representing how much weight is given to the specified amino acids on VH, VL (y axis) and RBD (x axis). Prediction results are evaluated by Precision-Recall curves of ACNN (Violet), Transformer (Gray), FCN (Red) and CNN (Blue).
  • b. The HCDR3 sequences of the predicted SARS-CoV-2 and Omicron variant binders are clustered by using an 80% sequence similarity. Cluster size represents the number of BCR sequences in the cluster. For each expanded cluster, the HCDR3 sequences are visualized as a sequence logo plot, where y-axis represents the frequency of the individual amino acid at the corresponding position in x-axis. The frequency of the dominating VH gene is listed above the logo.
  • c. Circos plot showing the frequency of antibodies encoded by the specified V region to J region pairing of the pan-SARS2 sequences.
  • d. The diversity of the four groups of BCR repertoire is analyzed, which is linked to the sample number of each group.
  • e. Binding of the predicted cross-reactive antibodies to RBD of SARS-CoV-2 and Omicron variants (left panel) was examined by ELISA. Representative OD reading is plotted as heatmap ranging from 0.05 to 5.0, and OD of 0.1 is used as cut-off value (n = 3 per group).
  • f. The SARS-CoV-2 Omicron variant (BA.1) pseudovirus neutralization curves of XBN-1, XBN-6 and XBN-11 mAbs were generated from luciferase readings at 8 dilutions (n = 3).
  • g. HCDR3 sequences of the XBN-1 and XBN-11 are aligned with the most convergent anti-SARS-CoV-2 antibodies from the published studies.
  • h. The HCDR3 sequence frequency of the dominant cluster (encoded by IGHV3-30 and IGKV1-13) of the pan-SARS group is shown.