MRICombo：一个基于深度学习的框架，用于跨异质性MRI的通用体积分割、分级分期和恶性检测

文献信息

中文摘要

肿瘤学中的全面磁共振成像（MRI）分析涉及多个相互关联的任务，包括体积分割、分级、分期和恶性检测。然而，大多数现有的深度学习模型是任务特异性或序列特异性的，缺乏异质性序列所需的泛化能力。在此，我们提出MRICombo，一个统一的多专家深度学习框架，用于跨9种异质性成像序列的通用解剖描绘和肿瘤表征。MRICombo使用来自2,354个个体的7,380个MRI序列开发，达到了最先进的性能，分割14个关键解剖结构的平均Dice相似系数为0.836，标记11种主要肿瘤类型的平均Dice相似系数为0.625。它在胶质瘤分级、膀胱癌和鼻咽癌分期以及乳腺和肝脏肿瘤恶性检测方面也达到了0.920的平均受试者工作特征曲线下面积（AUROC）。在四个独立数据集（来自734个个体的1082个序列）上的外部验证和迁移学习评估（512个个体）证实了稳健的跨协议泛化能力。此外，MRICombo支持缺失序列的灵活推理，并通过序列聚类和专家贡献分析提供决策可解释性。作为一个统一的临床解决方案，MRICombo显著降低了部署成本，并有可能简化诊断工作流程，支持更个性化的肿瘤护理。

英文摘要

Comprehensive magnetic resonance imaging (MRI) analysis in oncology involves multiple interrelated tasks including volumetric segmentation, grading, staging, and malignancy detection. However, most existing deep learning models are task-specific or sequence-specific, lacking the generalizability required for heterogeneous sequences. Here we present MRICombo, a unified multi-expert deep learning framework for universal anatomical delineation and tumor characterization across 9 heterogeneous imaging sequences. Developed using 7,380 MRI sequences from 2,354 individuals, MRICombo achieves state-of-the-art performance with mean Dice similarity coefficients of 0.836 for segmenting 14 critical anatomical structures and 0.625 for labeling 11 major tumor types. It also attains a mean area under the receiver operating characteristic curve (AUROC) of 0.920 for glioma grading, bladder and nasopharyngeal cancer staging, and breast and liver tumor malignancy detection. External validation on four independent datasets (1082 sequences from 734 individuals) and transfer learning evaluation (512 individuals) confirm robust cross-protocol generalizability. Furthermore, MRICombo supports flexible inference with missing sequences and offers decision interpretability through sequence clustering and expert contribution analysis. As a unified clinical solution, MRICombo significantly reduces deployment costs and has the potential to streamline diagnostic workflows, supporting more personalized oncology care.