A Digital Platform of an AI-aided pipeline for drug development in targeted therapy of cancer

Funded by Hangzhou 521 Talent programme, project launching in IIAT, Hangzhou, China.

• Detection and localization of cancer with morphological features and deep nets
• Genetic data processing using non-local graphical neural nets
• Multi-modality data alignment in feature space for target gene selection
• Online documents selection and semantic information extraction
• Simulation experiment for Molecular structure design

Deep learning multi-modal 3D registration

The deep learning based inverse-consistent registration model

• To the best of our knowledge, we are the first:
  • Deep learning based inverse-consistent multi-modality registration model that has been tested on 3D data (smaller network size than disentangled structure)
  • Deep learning model outperform the state-of-the-art and classic “SyN” method on difficult registration problems
  • SIMULTANEOUS joint training for cross-modal synthesis and registration

Paper accepted in CMIG (details will be given soon).

Cross-modality brain image synthesis using DicycleGAN

A new deformation-invariant GAN model (DicycleGAN), training from unpaired and unregistered images, output aligned data.

• Novel architecture based on deformable convolution and thin-plate-spline
• New alignment-based cycle-consistent loss
• A wide range of potential applications, e.g., synthesis of development of Alzheimer’s disease

C Wang, G Macnaught, G Papanastasiou, T MacGillivray, D Newby. Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. MICCAI-SASHIMI 2018 (oral)
C Wang et al. TMI 2019 (Under Review)
Two conference papers about improved model under review.

SaliencyGAN--Semi-supervised Salient Object Detection

The first semi-supervised salient object detection method

• New concatenated duo-GAN architecture
• With only 30% training data can have comparable performances with state-of-the-art fully supervised methods

C.Wang et al., IEEE TIP (under review)

Segmentation using multi-modality high dimensional 3D data

Two modified Unets concatenated for segmentation directly on high-resolution 3D data

• End-to-end and simultaneous training for segmentation and ROI-detection
• Can deal with high resolution 3D volume without resampling of input data

C. Wang, T. MacGillivray, G. Macnaught, G. Yang, D. Newby. A Two-Stage U-Net Model for 3D Multi-class Segmentation on Full-Resolution Cardiac Data. MICCAI-STACOM 2018.
Modality-merging deep seg net paper compared to domain-adaption based methods under review.

UFK-guided particle swarm optimizer for multi-modal brain image registration

A new PSO method designed for multi-modality image registration

• A single-point UKF is integrated in each iteration of PSO
• Lead to better and quick convergence for image registration problem

C Wang, K. Goatman, J. Boardman, E. Beveridge, D. Newby, S. Semple.Distance Oriented Particle Swarm Optimizer for Brain Image Registration. IEEE ACCESS. 2019.

MR-CT aortic and cardiac data registration algorithm and software

A new PSO method designed for multi-modality image registration

• A single-point UKF is integrated in each iteration of PSO
• Lead to better and quick convergence for image registration problem

C Wang, K Goatman, T MacGillivray, E Beveridge, Y Koutraki, J Boardman, C Stirrat, S Sparrow, E Moore, R Paraky, S Alam, M Dweck, C Chin, C Gray, D Newby, S Semple. Automatic multi-parametric MR registration method using mutual information based on adaptive asymmetric k-means binning. ISBI 2015.
C Wang et al. ESMRMB 2015.
C Wang et al. ISMRM 2014.
US Patent no. 9275432