基于跨模态掩码重建异常先验引导的 PET/CT 肿瘤分割
Anomaly-Guided Tumor Segmentation in PET/CT via Cross-Modality Masked Reconstruction
摘要
正电子发射断层扫描（PET）与计算机断层扫描（CT）的结合为肿瘤诊断提供了关键的代谢与解剖信息。针对全监督分割方法高度依赖高成本、高耗时的大规模精准标注数据的问题，研究提出一种基于异常检测引导的PET/CT肿瘤分割方法，旨在降低标注成本并提升自动化诊断水平。本文构建了一种基于双支路自动编码器网络的自监督异常检测框架。该模型仅利用正常PET/CT图像进行训练，以隐式学习正常生理代谢模式，并通过比较重建图像与原始图像的差异来识别异常病灶。在推理阶段，创新性地引入了基于标准摄取值（SUV）阈值的掩码策略，用以抑制疑似肿瘤区域对重建过程的干扰，提升重建质量。此外，将异常检测生成的概率图作为空间先验信息引入分割网络，引导模型聚焦于病变区域。实验结果表明，该方法在切片级和像素级评估中均显著优于现有的异常检测方法。该框架能有效利用异常先验知识提升肿瘤分割精度，在显著降低临床标注负担方面具有重要应用价值。
Abstract
The combination of positron emission tomography (PET) and computed tomography (CT) provides critical metabolic and anatomical information for tumor diagnosis. Addressing the issue that fully supervised segmentation methods highly rely on costly and time-consuming large-scale precisely annotated data, this study proposes an anomaly detection-guided PET/CT tumor segmentation method, aiming to reduce annotation costs and improve the level of automated diagnosis. This paper constructs a self-supervised anomaly detection framework based on a dual-branch autoencoder network. The model is trained solely on normal PET/CT images to implicitly learn normal physiological metabolic patterns, and identifies abnormal lesions by comparing the differences between the reconstructed images and the original images. During the inference stage, an SUV (Standardized Uptake Value) threshold-based masking strategy is innovatively introduced to suppress the interference of suspected tumor regions on the reconstruction process and improve reconstruction quality. Furthermore, the probability map generated by anomaly detection is incorporated into the segmentation network as spatial prior information, guiding the model to focus on the lesion areas. Experimental results demonstrate that the proposed method significantly outperforms existing anomaly detection methods in both image-level and pixel-level evaluations. This framework can effectively utilize anomaly prior knowledge to improve tumor segmentation accuracy, holding important application value in significantly reducing the clinical annotation burden.关键词
生物医学工程/肿瘤分割/肿瘤检测/自监督学习/医学图像重建Key words
Biomedical Engineering/Tumor Segmentation/Anomaly Detection/Self-Supervised Learning/Medical Image Reconstruction引用本文复制引用
冯志鸿,郭玙.基于跨模态掩码重建异常先验引导的 PET/CT 肿瘤分割[EB/OL].(2026-04-14)[2026-04-15].http://www.paper.edu.cn/releasepaper/content/202604-123.学科分类
肿瘤学/医学研究方法
评论