利用跨模型一致性作为结肠息肉分割在线可靠性信号

My Thoughts on Referee-Based Quality Estimation for Polyp Segmentation

Alright, let's break this down. In real-time colonoscopy, we're stuck because we don't have ground truth available during inference. So, these polyp segmentation models? They can fail silently, and we wouldn't even know. To address this, I've been working on Referee-Based Quality Estimation (RBQE). It's a reference-free framework designed to measure the agreement between a primary segmentation model and an independently trained "referee" model, both processing the same image. We've put this to the test on a substantial external benchmark – 1,223 images from four public datasets – using four different referee configurations. This setup lets me specifically look at referee independence and architectural diversity.

Using a straightforward "Agreement Dice" metric, a referee that's the same architecture as the primary model, but only differs in its random initialization, already provides a valuable reliability signal, with an impressive ROC-AUC of 0.923. That shows that independent training alone is a solid starting point. Now, cross-architecture referees take it further. The SegFormer-B0 model really shines here, achieving the best performance (ROC-AUC = 0.960). It's significantly better than the same-architecture control and UNet++, and beats a common Test-Time Augmentation baseline by a healthy 0.055 ROC-AUC, under identical protocols. Surprisingly, even with maximal architectural diversity, the prompt-coupled MedSAM referee doesn't perform as well.

Now, because cases where both models predict an empty mask (no polyp) are trivially separable, we've also run a restricted evaluation excluding these. The ROC-AUC drops to 0.876 for SegFormer-B0 (1,046 images) and 0.783 for the same-architecture control (975 images). But the interesting thing is that RBQE's advantage over the baselines actually increases on this subset. This is promising. Furthermore, as we progressively reject low-agreement predictions, RBQE helps increase the mean Dice of the remaining, retained predictions, which supports selective prediction. And, critically, this all only requires one additional deterministic referee forward pass during inference, so it doesn't add much computational burden. Ultimately, the work supports cross-model agreement as a practical and interpretable reliability framework for automated polyp segmentation, and I believe this has potential.

在实时结肠镜检查中，推理阶段缺乏真实标注，息肉分割模型可能会出现静默失败。为此，我们提出了基于裁判的质量估计（Referee-Based Quality Estimation, RBQE），这是一种无参考框架，用于衡量主分割模型与针对同一图像独立训练的“裁判”模型之间的一致性。我们在源自四个公开数据集、包含1,223张图像的标准化外部基准上对RBQE进行了评估，并采用四种裁判配置来分别探讨两个设计维度：裁判独立性与架构多样性。使用通用的“一致性Dice”描述符，仅在随机初始化上与主模型不同的同架构裁判便已能提供有价值的可靠性信号（ROC-AUC = 0.923），表明仅凭独立训练就足以奏效。跨架构裁判的表现进一步提升：SegFormer-B0取得了最佳性能（ROC-AUC = 0.960），显著优于同架构对照组和UNet++，并在相同评估协议下比代表性的测试时数据增强（TTA）基线高出0.055 ROC-AUC；相比之下，与提示关联的MedSAM裁判尽管具有最大的架构多样性，其表现却不及预期。鉴于空掩码一致性极易区分，我们还开展了排除此类样本的受限评估：ROC-AUC分别降至0.876（SegFormer-B0，1,046张图像）和0.783（同架构对照组，975张图像），但在此相同子集上，RBQE相对于两种基线的领先优势反而有所扩大。此外，随着低一致性样本被逐步剔除，RBQE提升了所保留预测结果的平均Dice值，有效支持了选择性预测，且在推理阶段仅需额外进行一次确定性的裁判模型前向传播。因此，本研究证实了跨模型一致性可作为自动化息肉分割任务中一种实用且具可解释性的可靠性评估框架。