Phantom-Anchored, Kv-Independent Deep Learning for Clinical Thoracic Iodine Map Generation on Photon-Counting CT
Abstract
Purpose
To develop a robust, generalizable iodine map generation framework for photon-counting CT (PCCT) using phantom-based calibration and patient-image fine-tuning.
Methods
The proposed 3D ResUNet combines U-Net localization with residual connections to enable deeper feature learning and reduce vanishing gradients. Five iodine rods (0–15mgI/cc) were placed in 20- and 40-cm abdominal phantoms and scanned on a PCCT (Siemens Alpha) at 120 kV using dual-/single-source modes and two dose levels. Vendor software generated VMIs at 50 and 70keV (inputs) and iodine maps (IMs; ground truth). Eight phantom datasets were split into six for training and two for validation; additional phantom VMIs acquired at 70, 90, and 140kV were used for independent testing. With IRB approval, eight pediatric cardiac CTA cases (120kV) were added for weakly supervised fine-tuning (VMI2IM-UNet-PP), and a phantom-only model (VMI2IM-UNet-P) was trained for comparison; two additional CTA cases were used for testing. Training used an NVIDIA Ada5000 GPU. Rod accuracy was assessed using mean absolute error (MAE), and patient IM similarity using structural similarity index (SSIM).
Results
Both VMI2IM-UNet-PP and VMI2IM-UNet-P outperformed least-squares (LS) fitting in phantom iodine accuracy, with no significant difference between the two models. For the newborn phantom at 70 kV, MAE was 1.581 mgI/cc (LS) vs 0.642 (PP) and 0.586 (P). For adult phantom at 140 kV, MAE was 1.728 (LS) vs 0.252 (PP) and 0.284 (P). For patients, VMI2IM-UNet-PP matched vendor contrast with lower noise and improved vessel detail relative to vendor IMs (SSIM 0.95–0.96), while VMI2IM-UNet-P showed mild bias and lower agreement (SSIM 0.86–0.90).
Conclusion
The proposed VMI2IM-UNet-PP, which anchors quantitative calibration in phantom data and leverages patient fine-tuning to learn the VMI-to-IM mapping, achieved improved iodine quantification accuracy and better preservation of patient vascular details. Importantly, the learned VMI-to-IM mapping generalized across PCCT acquisitions, supporting applications to all kV settings.