Interpretable Video Translation by generative AI:

• Created the largest dataset of talking head videos from YouTube.
• Established multi-person & lingual audio/video synchronization.
• Refined the facial landmark generation network for better articulation.
• Used diffusion to achieve immersive lip synchronization in videos with translated audio.

2D-3D style transfer for VR:

• Achieved real-time inference.
• Enabled human interaction for personalized customization.
• Preserved 3D visual reality.
• Utilized PyTorch3D & nvdiffrast as differential rendering.
• Preserved object style consistency by semantic style transfer.

Symmetric learning for Fracture Detection in Pelvic Trauma X-ray:

• Deployed in Chang Gung Memorial Hospital in Taiwan ans used by > 5000 patients.
• Paper accepted by ECCV 2020 with poster.
• Improved AUC from 0.95 to 0.98 and fracture recall from 0.89 to 0.93 (FPR=0.1).
• Mimicked radiologists to detect fractures by comparing bilateral symmetric regions.
• Focused on anatomical asymmetry with contrastive learning.

Deep Hierarchical Multi-label Classification of Chest X-ray:

• Paper accepted by MIDL 2019 with oral presentation.
• Paper accepted by Journal “Medical Image Analysis”.
• Mimicked radiologists to classify abnormality with clinical taxonomy.
• Improved classification AUC from 0.87 to 0.89.
• Robust to incompletely labeled data and preserved 85% performance drop.

Lung nodule detection in CT images:

• Achieved rank 6/2887 teams in the Skylake competition by Intel and Alibaba.
• Applied PyTorch, 3D UNet and Caffe, Faster RCNN in 1000 CT scans.
• Used fusion method to achieve false positive reduction.