通过训练中多个局部最优解，高效生成肿瘤分割不确定性图。
Rel-UNet: Reliable Tumor Segmentation via Uncertainty Quantification in nnU-Net
- 利用训练时多个局部最小值生成不确定性图
- 在KiTS23数据集上更快识别模糊区域，不确定性评分更低
- 无需改模型结构，适合各类分割模型快速部署
精准可靠的肿瘤分割对医学影像分析中的诊断、治疗规划和疗效监测至关重要。然而，现有分割模型往往缺乏预测不确定性量化机制，影响临床决策可信度。本研究提出一种基于深度学习的肾肿瘤分割不确定性量化新方法，利用训练过程中的多个局部最小值生成不确定性图，无需修改原始模型架构且计算开销低。在KiTS23数据集上的评估表明，该方法能更快速地识别模糊区域，且不确定性评分低于以往方法。生成的不确定性图可揭示模型置信度，显著提升分割可靠性，助力更准确的医疗诊断。该方法计算高效、模型无关，无需架构改动，可广泛适配各类分割模型。
原文摘要 · Abstract (English)
Accurate and reliable tumor segmentation is essential in medical imaging analysis for improving diagnosis, treatment planning, and monitoring. However, existing segmentation models often lack robust mechanisms for quantifying the uncertainty associated with their predictions, which is essential for informed clinical decision-making. This study presents a novel approach for uncertainty quantification in kidney tumor segmentation using deep learning, specifically leveraging multiple local minima during training. Our method generates uncertainty maps without modifying the original model architecture or requiring extensive computational resources. We evaluated our approach on the KiTS23 dataset, where our approach effectively identified ambiguous regions faster and with lower uncertainty scores in contrast to previous approaches. The generated uncertainty maps provide critical insights into model confidence, ultimately enhancing the reliability of the segmentation with the potential to support more accurate medical diagnoses. The computational efficiency and model-agnostic design of the proposed approach allows adaptation without architectural changes, enabling use across various segmentation models.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。