Аңдатпа
Accurate lymph node assessment on computed tomography (CT) is essential for cancer staging and treatment planning, yet automation remains challenging due to small node size, low soft-tissue contrast, and extreme class imbalance (foreground voxels comprising below 1% of the volume). We propose the Cross-Slice Transformer (CST), a lightweight 2.5D architecture that learns adaptive slice importance through statistical descriptor extraction and transformer-based attention over five consecutive CT slices. CST is evaluated against three 2.5D baselines (U-Net, TransUNet, SwinUNet) and the 3D nnUNet framework on two publicly available TCIA datasets: a multi-region dataset (Ver1, 296 patients, mediastinal and abdominal) and a mediastinal-focused dataset (Ver2, 206 patients). nnUNet-3D achieves the highest overall Dice on both datasets (0.589 on Ver2), confirming that full volumetric context remains the strongest approach when GPU resources permit. Our central research questions instead concern 2.5D architectural design and dataset composition, which a single 3D model cannot address. On Ver2, CST achieves the highest Dice among 2.5D models ( $0.543 \pm 0.044$ ), with the highest recall among 2.5D models ( $0.618 \pm 0.024$ ) and a compact footprint of 7.97 M parameters, substantially fewer than the 2.5D baselines (25–99 M). Relative to nnUNet-3D, the advantages of CST lie in this compactness and in node-level detection rather than in overall Dice. Node-level detection analysis reveals that CST leads all models including nnUNet-3D on detection rate across all size strata on Ver2: 53.6% for small nodes (< 10 mm), 82.9% for medium (10–15 mm), and 92.5% for large ( $\geq 15$ mm) at IoU $\geq $ 0.1. All models improve consistently when trained on anatomically focused single-region data, with Dice gains of 0.047–0.152 points from Ver1 to Ver2, demonstrating that dataset composition is a primary driver of segmentation performance across all architectural families. Ablation experiments confirm the contribution of each CST component: adaptive transformer attention (+ 0.006 Dice over fixed uniform weights), the interaction descriptor (+ 0.008 over a reduced descriptor), and per-case intensity normalisation (+ 0.008 over fixed HU windowing). These results provide evidence-based guidance for 2.5D architecture selection and dataset curation in clinical lymph node segmentation.