Deep Learning CT Harmonization

Developing physics-informed deep learning models to generate scanner-agnostic, quantitative mono-energetic images from multi-energy photon-counting CT images.

Multi-energy CT Image Data Acquisition

Digital human and cylindrical phantoms were virtually scanned using CdTe-, CZT-, and Si-based photon-counting CT systems to generate quad-energy CT datasets across a range of acquisition settings. These synthetic datasets were then used to train, validate, and test the deep-learning mono-energetic image synthesis model.

Deep Learning Framework

Problem – Higher Cross-Scanner Variability

Pre-harmonized quad-energy CT images, demonstrating the variation in mean pixel values (CT numbers) for Ca (100 mg/mL), I (15 mg/mL), and Gd (10 mg/mL) inserts across four energy bins (rows) and CT scanners (columns). All images were acquired under identical acquisition settings.

Pre-harmonization: coefficient of variation in % for mean CT numbers (y-axes) at different energy bins (x-axes) measured across various CT scanners (top-left), phantom sizes (top-right), reconstruction kernels (bottom-left), and exposure levels (bottom-right), while for each keeping other factors constant. CT-number variability across inserts ranged from 0.0% to 45.7%, indicating that a single algorithm cannot reliably generate mono-energetic images across all conditions without condition-specific calibration.

Reconstructed 35 – 85 keV Mono-Energetic Images

Ground-truth References

Ground-truth
(20-cm cylindrical phantom)

Ground-truth
(30-cm cylindrical phantom)

Conventional Physics-based Mono-Energetic Images

Deep Learning-based Mono-Energetic Images

Iodine Maps Generated Using Mono-Energetic Images

Conclusion

By integrating realistic, physics-informed in silico simulations with a conditional U-Net architecture, the model learns spectral-spatial relationships that allow accurate and continuous mono-energetic image synthesis across energy levels. DL-synthesized mono-energetic images show strong quantitative agreement and structural fidelity with ground-truth data, maintaining inter- and intra-scanner consistencies.

This study establishes the feasibility of using AI-based approaches, acquainted with in silico imaging frameworks that provide ground-truth mapped datasets otherwise impractical experimentally, for cross- scanner spectral harmonization – enabling standardized, scanner-agnostic imaging to facilitate quantitative CT.

Want to learn more?

← Back

Mridul Bhattarai, PhD