Deep learning based processing for quantitative myocardial perfusion
Deep learning based processing for quantitative myocardial perfusion MRI Cian M. Scannell, Mitko Veta, Adriana D. M. Villa, Eva Sammut, Jack Lee, Marcel Breeuwer and Amedeo Chiribiri
Myocardial perfusion quantification Arterial Input Function Myocardial Tissue Curve
Myocardial perfusion quantification Arterial Input Function 2 CXM Myocardial Tissue Curve
An automated pipeline
An automated pipeline
Peak LV enhancement detection
Peak LV enhancement detection
Bounding box
Bounding box
Segmentation Motion correction Scannell et al. IEEE TMI 2019
Segmentation
RV insertion points
RV insertion points Xu et al. Supervised Action Classifier: Approaching Landmark Detection as Image Partitioning. MICCAI 2017
An automated pipeline Quantification using Bayesian inference Scannell et al. ar. Xiv: 1906. 02540 2019 (Submitted to Medical Image Analysis)
Training details • 175 patients (1050 individual slices) • Split into 135/10/30 (train/validation/test) Trained with • Data augmentation • Batch normalisation • Early stopping • Dropout Test time augmentation used for inference
Results
Results • Peak LV enhancement detection - Mean (standard deviation) error of 1. 49 (1. 40) time frames • Bounding box detection - Mean (standard deviation) Dice similarity of 0. 93 (0. 03) • Myocardial segmentation - Mean (standard deviation) Dice similarity of 0. 80 (0. 07) • RV insertion point detection - Mean (standard deviation) Euclidean distance between points of 2. 28 (1. 8) mm
Results
AHA bullseye representation
Quantitative MBF results • Intra-class correlation coefficient of 0. 867 [0. 829, 0. 896].
Thank you cian. scannell@kcl. ac. uk @cianmscannell
- Slides: 21