Neural Network Approach for Photon-counting Detection – The First Step: PPE Correction Ruibin Feng, Ph.D. Ge Wang, Ph.D. David Rundle Biomedical Imaging Center, Biomedical Imaging Center, JairiNovus Technologies Ltd. CBIS/BME, RPI CBIS/BME, RPI david.rundle@jairinovus.com fengr@rpi.edu wangg@rpi.edu Nov. 19, 2017
Outline • X-ray Detectors: EIDs vs. PCDs • PCD Data Degradation • Trigger Threshold Correction • Monte-Carlo Simulation • Discussions & Conclusion
Energy-integrating Detectors (EIDs) • Mature technology in all current x-ray scanners • Energy integration over the entire x-ray spectrum
Drawbacks of EIDs • Energy-dependent information lost – Linear attenuation not tissue-type sensitive • Data quality degenerated due to the dark current (electric/Swank noise) – Low SNR • Low-energy photons under weighted – Poor contrast, beam-hardening
Photon-counting Detectors (PCDs) • Voltage cross the threshold counted, individually and energy-sensitively • Multiple energy windows spanning the spectral dimension for CT imaging
Advantages of PCDs • Spectrally unique contrast – K-edge and fluorescence imaging, beam-hardening avoidance • Low radiation dose – No electronic noise, balanced photon weights, improved SNR • High spatial resolution – Desirable for radiomics
PCD Data Degradation • Pulse Pileup Effect (PPE) • Charge sharing • K-escape x-rays • Compton scattering
Pulse Pileup Effect (PPE) Missed Counts
Pulse Pileup Effect (PPE) Distorted Energy
Pulse Pileup Effect (PPE) • PCDs degrade in the performance of detection tasks when the count rate exceeds 20% of the maximum rate • Current compensation/calibration methods are not optimal and difficult to extend for different applications – Model must be accurate to describe the detection process – Optimization must be specific to address intended tasks such as material decomposition or contrast estimation
NN-based Trigger Threshold Correction True Measurement, DLR = 4.07% True Measurement, DLR = 23.21% True Measurement, DLR = 56.03% 250 1400 6000 1200 5000 200 1000 4000 150 800 Count Count Count 3000 600 100 2000 400 50 1000 200 0 0 0 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 Energy Bin Energy Bin Energy Bin Distorted Measurement, DLR = 4.07% Distorted Measurement, DLR = 56.03% Distorted Measurement, DLR = 23.21% 200 2000 1000 180 1800 900 160 1600 800 140 1400 700 120 1200 600 Count Count Count 100 500 1000 80 400 800 300 60 600 40 200 400 100 20 200 0 0 0 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 Energy Bin Energy Bin Energy Bin
Trigger Threshold • X-ray tube energy: 120 KeV • Normal threshold: < 120 KeV • Tigger threshold: > 120 KeV Signal strength over the trigger threshold indicates whether PPE occurs and how severe it is
NN-based Correction for PPE Normal Thresholds NN Trigger Thresholds Distorted Measurement Corrected Measurement
Monte-Carlo Simulation Normalized Transmitted Spectrum 0.045 • X-ray spectrum 0.04 – TASMICS 0.035 • 43 Combinations of 0.03 Attenuators Probablity 0.025 – Water, Bone, Blood w. 20% Gd 0.02 – Thickness T = {20, 30} cm – Bone: T(bone) = {0, 1, 3, 5} cm 0.015 – 20% Gd: T(Gd) = [0:4:20] cm 0.01 – T(water) = T - T(bone) - T(Gd) 0.005 0 0 50 100 150 Energy
Monte-Carlo Simulation Normailzed Pulse Shape (deadtime = 1 s) Unipolar Pulse 1 • Pulse Shaper Bipolar Pulse 0.8 Unipolar Pulse Energy (keV) 0.6 Bipolar Pulse 0.4 0.2 0 0 1 2 3 4 5 6 7 8 9 Time ( s) • Detector Type Paralyzable Nonparalyzable
Monte-Carlo Simulation Training and Testing Datasets • 1,000 measurements for each attenuator • Dataset 1: – 36 attenuators – Training, validation, testing = 60%, 20%, 20% • Dataset 2: – 7 attenuators
Monte-Carlo Simulation • Deadtime Loss Ratio (DLR) Paralyzable detector: Nonparalyzable detector: • Coefficient of Variation (COV)
Monte-Carlo Simulation • Neural Network Model Fully-connected NN with 1 hidden layer 512 hidden units Dropout and L2 regularizer • Unbiased Estimator
Numerical Results • Unipolar Pulse & Paralyzable Detector Comparison with Detector Measurement Comparison with Unbiased Estimator 0.4 0.05 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 1 0.35 Proposed Corrector: Dataset 2 Proposed Corrector: Dataset 2 Detector Measurement: Dataset 1 Unbiased Estimator: Dataset 1 0.04 0.3 Detector Measurement: Dataset 2 Unbiased Estimator: Dataset 2 0.25 0.03 COV COV 0.2 0.02 0.15 0.1 0.01 0.05 0 0 10 0 10 1 10 0 10 1 DLR (%) DLR (%)
Numerical Results • Bipolar Pulse & Paralyzable Detector Comparison with Detector Measurement Comparison with Unbiased Estimator 0.35 0.05 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 2 0.3 Proposed Corrector: Dataset 2 Detector Measurement: Dataset 1 Unbiased Estimator: Dataset 1 0.04 Detector Measurement: Dataset 2 Unbiased Estimator: Dataset 2 0.25 0.03 0.2 COV COV 0.15 0.02 0.1 0.01 0.05 0 0 10 0 10 1 10 0 10 1 DLR (%) DLR (%)
Numerical Results • Unipolar Pulse & Nonparalyzable Detector Comparison with Detector Measurement Comparison with Unbiased Estimator 0.4 0.05 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 1 0.35 Proposed Corrector: Dataset 2 Proposed Corrector: Dataset 2 Detector Measurement: Dataset 1 Unbiased Estimator: Dataset 1 0.04 0.3 Detector Measurement: Dataset 2 Unbiased Estimator: Dataset 2 0.25 0.03 COV COV 0.2 0.02 0.15 0.1 0.01 0.05 0 0 10 0 10 1 10 0 10 1 DLR (%) DLR (%)
Numerical Results • Bipolar Pulse & Nonparalyzable Detector Comparison with Detector Measurement Comparison with Unbiased Estimator 0.35 0.05 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 1 Proposed Corrector: Dataset 2 Proposed Corrector: Dataset 2 0.3 Detector Measurement: Dataset 1 Unbiased Estimator: Dataset 1 0.04 Detector Measurement: Dataset 2 Unbiased Estimator: Dataset 2 0.25 0.03 0.2 COV COV 0.15 0.02 0.1 0.01 0.05 0 0 10 0 10 1 10 0 10 1 DLR (%) DLR (%)
Future Plan for PPE Correction • Systematic Simulation Study • Phantom Experiments • Preclinical Testing How to Collect Unbiased Data? – Perform realistic simulation with professional software tools – Reduce the incident flux for PPE-free data via time integration
Future Plan for CS Correction Charge Sharing: one photon is detected by multiple pixels with lower energies
Conclusion We have proposed an NN/ML approach to handle PPE and other artifacts in PCD data • Extract an optimal relationship between PCD data before and after degradation of any kind • Potentially, the NN/ML approach can outperform the existing patented methods for PCD data correction, and improve photon-counting CT image reconstruction
Recommend
More recommend
