Head-and-Neck PET/CT Lesion Segmentation via SSIMH and SegResNet

Shuai Huang, Yuqi Cheng, Shilin Yu, Hongtu Zhu

Keywords: HECKTOR25 · Head and neck cancer · PET/CT segmentation · Domain adaptation · SegResNet

Abstract: We present a simple and effective pipeline for automatic detection and segmentation of primary tumors and lymph nodes in FDG-PET/CT for HECKTOR 2025 Task 1. The method starts with an anatomy-aware pre-crop of the head-and-neck region to suppress irrelevant context, followed by modality-specific intensity normalization with soft clamping. To mitigate cross-center domain shift, we apply single-subject, SSIM-guided histogram matching without external references. For segmentation, we use a residual U-Net–style SegResNet with deep supervision and a combined Dice + Cross-Entropy loss. Training employs stratified five-fold cross-validation with foreground-centered sampling to emphasize small lesions. At inference, we use sliding-window tiling on the cropped volumes and straightforward post-processing to remove small isolated components. We ensemble the five cross-validation models by averaging per-voxel logits before softmax, which improves stability across centers and lesion types with moderate computational overhead. The overall recipe is concise and reproducible while remaining competitive. Our contributions are: (i) a compact, anatomy-aware preprocessing scheme for head-and-neck PET/CT; (ii) an SSIM-guided, per-subject histogram matching strategy for robust domain adaptation; and (iii) a strong SegResNet baseline with a simple training/inference protocol augmented by a lightweight five-model ensemble. (Team name: BIGS2)

Submission Number: 14