Protocol-Aware Multi-Slice Deep Learning Framework for Resolution Enhancement In Conventional Coronary CT Angiography Using Photon-Counting Detector CT As Reference
Abstract
Purpose
To enhance spatial resolution in all three dimensions (x, y, z) of coronary CT angiography (cCTA) acquired with conventional energy-integrating detector (EID)-based CT systems, using a protocol-aware multi-slice deep learning model trained with resolution-matched data from photon-counting detector (PCD) CT.
Methods
A protocol-aware training strategy was developed using PCD-CT data to mimic the spatial resolution of EID-CT, based on reconstruction kernel information (e.g., Br36) extracted from clinical protocols. A matched-kernel PCD-CT reconstruction (e.g., Br36) served as the network input, while a paired sharper-kernel high-resolution (HR) reconstruction (e.g., Bv72) was used as the training label. For network training, eight coronary CT angiography (cCTA) cases were acquired on a PCD-CT scanner (NAEOTOM Alpha, Siemens Healthineers) using ultra-high-resolution (UHR) mode. From each case, 128×128 axial image patches were extracted from matched-kernel images using three adjacent slices as input channels, with the corresponding central slice from the HR reconstruction used as the label. This multi-slice input design preserved spatial context along the z-direction. The network architecture was a multi-slice U-Net with nine modules, each consisting of convolutional layers, batch normalization, and exponential linear unit (ELU) activations. Down-sampling and up-sampling were implemented using max pooling and transposed convolutions, respectively, with skip connections to retain spatial detail.
Results
The proposed model enhanced resolution across all spatial dimensions. Coronal and sagittal reformats showed improved through-plane sharpness and vessel continuity, particularly along the z-axis. In axial views, calcified plaque and coronary artery borders appeared sharper with reduced blooming artifacts. Line profile analysis across coronary calcification revealed steeper intensity transitions and an increased lumen diameter in the enhanced network output image, indicating improved spatial resolution.
Conclusion
A protocol-aware multi-slice deep learning model trained on kernel-matched PCD-CT data effectively enhances spatial resolution in all three dimensions for conventional EID-based coronary CT angiography.