Proof of Concept Study of Principal Component Analysis for Quantitative Digital Subtraction Angiography (qDSA) Blood Velocity Estimation
Abstract
Purpose
To advance the qDSA algorithm beyond mean velocity estimations by quantifying systolic and diastolic waveform parameters using 2D image-derived features.
Methods
A reduced order model of fixed-rate (2.0 mL/s) iodine contrast injection into pulsatile arterial flow was used to simulate high frame rate DSA (30 fps). X-ray projections were converted into Time-Attenuation Maps (TAMs) to visualize intensity versus vessel distance and time in 2D. 20,000 TAMs for a 20 cm long vessel were generated with varying cardiac waveforms. The waveforms had a peak systole (PS) velocity range of [10, 200] cm/s, end diastole (ED) range of [1, 100] cm/s, systolic duration ratio (Tau) [0.2, 0.9], and heart rate [30 120] bpm. Principal component analysis (PCA) was performed on the training set to extract the first 100 mode images and weights of TAMs. PCA weights were inputs for a multilayer perceptron with outputs of PS, ED, and Tau. Curriculum learning trained the network on smaller length TAMs. K-fold validation was performed with 5 folds (3-fold train, validate on 4th, test on 5th, repeat for 5 runs). After training, accuracy versus SNR was evaluated on TAMs with additive gaussian noise.
Results
Time to estimate velocity parameters from a TAM was 10.69+/-1.19 ms. For full length TAMs, mean biases were 0.24 cm/s for PS, -0.27 cm/s for ED, and -1.17e-3 for Tau. The 95% limits of agreement were [-13.00, 13.48] cm/s for PS, [-6.02, 5.49] cm/s for ED, and [-0.05, 0.05] for Tau. For TAMs greater than 5 cm in length, mean absolute errors were below 7.8 cm/s for PS, 3.2 cm/s for ED, and 0.019 for Tau. An SNR of 8 was required to obtain estimations within 10% of noiseless TAMs.
Conclusion
A proof-of-concept method for quantifying blood velocities at systole and diastole from high-fps 2D images was developed and evaluated.