Image Translation by Latent Union of Subspaces for Cross-Domain - - PowerPoint PPT Presentation

Image Translation by Latent Union of Subspaces for Cross-Domain Plaque Detection

Na Natio ional I l Instit itutes o

• f H

Healt lth Cl Clinical Ce Center

Yingying Zhu1, Daniel C. Delton1, Sungwon Lee1, Perry J. Pickhardt2, Ronald M. Summers1

1 Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, National Institutes of Health Clinical Center, Bethesda, MD 20892, USA 2 School of Medicine and Public Health, University of Wisconsin, Madison, WI 53706, US

Challenges

• Generalization of calcified plaque detection model on

different domains

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

No Non-Co Contrast CT CT (NCCT CCT) Ha Hand La Labelled Pl Plaque ue Calcified plaque Co Contrast Enhanced CT CT (CE CECT CT) Calcified plaque

Can we translate image across domains?

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

Current Cross-Domain Image Translation Method

Co Contrast-En Enhanced d CT Sy Synt nthet hetic No Non-Contrast CT (Li Liu et al. 2018) Calcified plaque Calcified plaque

Ca Calci cified plaques are not

• t preserve

ved after im image transla latio ion

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

Extract Small Patches to Union of Subspaces

Calcified plaque

Di Different image patch lies in different clusters

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

Cross-Domain Image Translation by Shared Union of Subspaces

Contrast(Enhanced(CT Sy Synt nthet hetic Non0Contrast(CT((our(model)

Ca Calci cified plaques are preserve ved much ch be better a after i image ge t translation

Results

• Contrast Enhanced CT (CECT) calcified plaque detection &

segmentation by Mask-RCNN

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

He,%Kaiming et%al.%“Mask%R4CNN.” 2017%IEEE%International%Conference%on%Computer%Vision%(ICCV) (2017):%298042988.

Tr Training Data No Non-Con Contrast CT CT (N (NCCT CCT) No Non-Con Contrast CT CT (N (NCCT CCT) No Non-Con Contrast CT CT (N (NCCT CCT) Te Testing data Synthetic NCCT by Cycle GANS Zhu et. al Synthetic NCCT by UNIT Liu et. al Synthetic NCCT by our model Precision 60.5±2.87% 63.2±2.64% 77.5 77.5±2.58 2.58% Recall 65.7±3.21% 69.5±3.05% 78.6 78.6±2.87 2.87% Dice 0.534±0.236 0.566±0.198 0.676 0.676±0.176 0.176

Thanks

Na Natio ional I l Instit itutes o

• f H

Healt lth Cl Clinical Ce Center

Na Nati tional I Insti titu tute tes o

• f H

f Health th Clin Clinic ical l Ce Center

Email:'yingying.zhu@nih.gov