Voxel-Wise Deep Learning Segmentation of Hydroxyapatite and Iodine in Spectral Photon-Counting C

09 Apr 20264 min read0 viewsJournal Feed

GIST

by Nadine Francis, Mohamed L. Seghier, Nabil Maalej, Aamir Raja Accurate non-invasive identification of hydroxyapatite (HA) deposits is important for diagnosing calcific musculoskeletal disease and quantifying vascular calcification, but conventional and dual-energy CT often struggle to distinguish HA from iodinated contrast because of overlapping attenuation, noise, and beam-hardening artifacts.

Spectral photon-counting CT (SPCCT) offers improved energy resolution and spatial fidelity, yet most deep-learning approaches in spectral CT focus on continuous density regression or anatomical segmentation rather than direct voxel-wise material labeling. We developed SPFF–UNet, a spectral-preserving 3D segmentation model for direct classification of HA and iodine concentrations from five-bin SPCCT volumes without material-decomposition preprocessing.

A cylindrical phantom containing twelve materials was scanned at 0.1 mm isotropic resolution, including five HA concentrations, three iodine concentrations, three soft-tissue equivalents, and water. SPFF–UNet integrates spectral squeeze-excitation, EnergyFiLM, and FourierGate to preserve and exploit multi-energy information throughout the network.

The model was trained for thirteen-class voxel-wise segmentation and compared with five established 3D architectures under matched training conditions.

Clinical Editorial

Spectral Preservation in Voxels: Study Design and Purpose

This phantom study targets non-invasive identification of hydroxyapatite (HA) and iodine using spectral photon-counting CT (SPCCT), addressing limitations of conventional and dual-energy CT in distinguishing overlapping attenuation, noise, and beam-hardening artifacts.
The objective was to develop a voxel-wise segmentation model that directly classifies HA and iodine concentrations without material-decomposition preprocessing.

Study Design and Technical Approach

A cylindrical phantom containing twelve materials was scanned at 0.1 mm isotropic resolution, including five HA concentrations, three iodine concentrations, three soft-tissue equivalents, and water.
The authors introduced SPFF–UNet, a spectral-preserving 3D segmentation network designed for direct concentration-aware classification of HA and iodine across five SPCCT energy bins.
SPFF–UNet incorporates spectral squeeze-excitation, EnergyFiLM, and FourierGate components to maintain and leverage multi-energy information throughout the network.
The model was trained for thirteen-class voxel-wise segmentation and benchmarked against five established 3D architectures under matched training conditions.

Performance and Comparative Context

On a held-out phantom scan, SPFF–UNet achieved the highest macro-averaged metrics among the tested architectures: Dice 0.72, IoU 0.59, sensitivity 0.73, and precision 0.71.
The strongest comparator, ResUNet++, trailed with Dice 0.66, IoU 0.46, sensitivity 0.67, and precision 0.61.
Gains were most pronounced for mid/low-contrast HA and low-concentration iodine, accompanied by reduced slice-to-slice variability and a lower rate of HA–iodine misclassifications.

Interpretation and Implications

The study supports the concept that preserving spectral information and employing targeted spectral modulation can enhance concentration-aware voxel classification from SPCCT data.
This work represents a phantom-based proof-of-concept and explicitly frames its findings as preliminary, suitable for guiding future in vivo validation.

Limitations and Open Questions

The evidence is restricted to a single phantom dataset for held-out evaluation; generalizability to in vivo imaging remains untested.
No clinical outcome data or patient-centered metrics are presented; applicability to diagnostic workflows is not established within this report.