"Multi-Material SPCCT Phantom Dataset with Five Energy Bins: Hydroxyapatite (50\u2013800 mg\/cm\u00b3), Iodine (5\u201315 mg\/mL), and Soft-Tissue"

"Spectral photon-counting CT (SPCCT) enables material segmentation from multi-energy measurements, but many conventional 3D segmentation backbones downsample the energy-bin axis, weakening the spectral cues needed to separate hydroxyapatite (HA) from iodine and to resolve discrete concentrations. We propose S\\textsuperscript{3}UNet, a spectral--spatial 3D U-Net that preserves the spectral axis via anisotropic pooling\/upsampling and integrates spectral-aware attention, a Swin-inspired channel-mixing bottleneck, and a refinement head. On a SPCCT phantom dataset (five scans; five energy bins; HA50--HA800, iodine I5--I15, soft-tissue equivalents, and water), trained on four scans and evaluated on a held-out external scan, S\\textsuperscript{3}UNet achieves $0.84\\pm0.04$ foreground macro-averaged Dice and $0.86\\pm0.03$ foreground macro-averaged precision, outperforming 3D U-Net ($0.48\\pm0.07$; $0.49\\pm0.09$) and SwinUNETR ($0.65\\pm0.02$; $0.55\\pm0.01$). Single-factor ablations show the largest degradations in macro-averaged performance for IsoPool and NoRefine. Finally, segmentation-derived rod-wise readout achieves the lowest pooled calibration MAE (HA+I) without a dedicated regression head. These results support spectral-axis-preserving architectures for robust SPCCT material labeling and segmentation-driven quantification."

光谱光子计数计算机断层扫描（Spectral photon-counting CT, SPCCT）可基于多能量测量实现材料分割，但多数传统三维分割骨干网络会对能量仓轴进行下采样，削弱了区分羟基磷灰石（hydroxyapatite, HA）与碘、分辨离散浓度所需的光谱线索。本文提出S³UNet——一种光谱-空间三维U-Net，其通过各向异性池化/上采样保留光谱轴，并集成了光谱感知注意力机制、受Swin架构启发的通道混合瓶颈模块与细化头。在一款光谱光子计数CT体模数据集（包含5次扫描、5个能量仓，涵盖浓度范围为HA50–HA800的羟基磷灰石、I5–I15的碘、软组织等效材料与水）中，我们以4次扫描作为训练集，以预留的外部扫描作为测试集。实验结果显示，S³UNet实现了0.84±0.04的前景宏平均Dice系数与0.86±0.03的前景宏平均精确率，性能优于3D U-Net（0.48±0.07；0.49±0.09）与SwinUNETR（0.65±0.02；0.55±0.01）。单因子消融实验表明，采用IsoPool（各向同性池化）与NoRefine（无细化头）配置时，模型的宏平均性能出现最大幅度的衰减。最终，无需专用回归头的分割驱动逐棒读出策略，实现了最低的合并校准平均绝对误差（HA+I）。上述结果证实，保留光谱轴的架构可实现鲁棒的光谱光子计数CT材料标注与分割驱动的量化任务。