RC-aware nnU-Netv2：基于多模态MRI的预处理与治疗后胶质瘤分割
RC-aware nnU-Netv2 for Pre-treatment and Post-treatment Glioma Segmentation Using Multimodal MRI
- School of Statistics, East China Normal University（华东师范大学统计学院）
- Department of Biostatistics, School of Global Public Health, New York University（纽约大学全球公共卫生学院生物统计学系）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
提出RC-aware nnU-Netv2框架，通过RC感知联合目标与治疗状态路由，解决预处理和治疗后胶质瘤分割中RC目标不一致问题，在BraTS-GLI 2025中排名第二。
AI中文摘要：
BraTS 2025 Lighthouse Challenge任务1（BraTS-GLI 2025）评估了预处理和治疗后多模态MRI中的胶质瘤分割。切除腔（RC）仅适用于治疗后病例，导致两个队列的目标定义不同。我们开发了RC-aware nnU-Netv2，这是一个将病灶和边界感知的单队列训练与RC感知联合目标及治疗状态引导路由相结合的框架。在池化训练期间，联合目标对所有四个区域通道应用标准及边界加权的二元交叉熵，同时使用标准Dice监督当前小批量中具有非空目标的通道，并对整个小批量中为空的目标通道施加受控的假阳性惩罚。这种空目标处理对于预处理病例中的全零RC目标尤为重要。除了标准nnU-Netv2增强流程外，本提交不使用合成肿瘤生成、实时GliGAN增强、模型级概率平均、投票、多折融合或多架构集成。每个病例被路由到恰好一个专用模型。我们的提交在BraTS-GLI 2025中排名第二。在官方盲测集上，我们的方法在增强肿瘤、RC、肿瘤核心和全肿瘤上分别实现了平均病灶级Dice分数0.7878、0.8709、0.7923和0.8715，以及平均NSD@1.0分数0.8309、0.8712、0.7980和0.8336。在配对治疗后分析中，与联合基线训练相比，RC感知联合模型将平均病灶级Dice提高了0.042（95%置信区间：[0.028, 0.057]），平均NSD@1.0提高了0.043（95%置信区间：[0.028, 0.059]）。
英文摘要：
BraTS 2025 Lighthouse Challenge Task 1 (BraTS-GLI 2025) evaluates glioma segmentation in pre-treatment and post-treatment multimodal MRI. The resection cavity (RC) is applicable only to post-treatment cases, creating different target definitions across the two cohorts. We developed RC-aware nnU-Netv2, a framework that combines lesion- and boundary-aware single-cohort training with an RC-aware joint objective and treatment-status-guided routing. During pooled training, the joint objective applies both standard and boundary-weighted binary cross-entropy to all four region channels, while using standard Dice supervision for channels with a non-empty target in the current mini-batch and a controlled false-positive penalty for channels that are empty across the mini-batch. This empty-target handling is particularly relevant to the all-zero RC target in pre-treatment cases. Beyond the standard nnU-Netv2 augmentation pipeline, the submission does not use synthetic tumor generation, on-the-fly GliGAN augmentation, model-level probability averaging, voting, multi-fold fusion, or multi-architecture ensembling. Each case is routed to exactly one specialized model. Our submission ranked second in BraTS-GLI 2025. On the official blind test set, our method achieved mean lesion-wise Dice scores of 0.7878, 0.8709, 0.7923, and 0.8715 and mean NSD@1.0 scores of 0.8309, 0.8712, 0.7980, and 0.8336 for enhancing tumor, RC, tumor core, and whole tumor, respectively. In paired post-treatment analysis, the RC-aware joint model improved mean lesion-wise Dice by 0.042 (95% CI: [0.028, 0.057]) and mean NSD@1.0 by 0.043 (95% CI: [0.028, 0.059]) compared with joint baseline training.
