DCPFC dataset

 Brain parcellation provides fundamental insights into the organizational principles of the human brain. Recently, there has been an increasing interest in fMRI connectivity-based parcellation to non-invasively delineate functional brain subdivisions. However, current clustering methods struggle to fully exploit the complex and nonlinear features of the brain, compromising the reproducibility and spatial continuity of the resulting parcels. Moreover, the derived parcels are typically labeled based on anatomical axes and lack detailed functional characterization, limiting a deeper understanding of brain function. To address these issues, we propose a Deep Clustering Parcellation with Functional Characterization (DCPFC) framework and apply it to the striatal subdivisions of individuals or populations in two open-access resting-state fMRI datasets. This framework introduces a deep clustering network to extract and reconstruct nonlinear features while simultaneously performing clustering through iterative optimization. An additional spatial constraint with manifold distance is proposed to ensure the spatial continuity of parcels. Importantly, the functional characteristics of these striatal parcels are further decoded by mapping the reconstructed cluster centroids to the cerebral cortex. Compared with other prevalent connectivity-based parcellation (CBP) methods, the proposed framework demonstrated superior reproducibility and spatial continuity, as well as comparable functional homogeneity. Furthermore, the striatum is divided into 6 parcels, delineated based on somatomotor, attention (dorsal and ventral), and frontoparietal networks. Significant correlations were observed between cognitive flexibility and delay discounting with the size of these parcels, highlighting the efficacy of our framework in capturing the cognitive-related functional topography of the brain. The functionally explainable striatal subdivisions generated by this framework further enhance our understanding of the brain and provide a solid foundation for advancing clinical research.

大脑分区（brain parcellation）可为揭示人脑的组织架构原则提供基础性见解。近年来，基于功能磁共振成像（fMRI）连接组的大脑分区技术愈发受到学界关注，旨在无创勾勒脑功能亚区。然而，现有聚类方法难以充分挖掘大脑复杂且非线性的特征，致使所得分区的可复现性与空间连续性大打折扣。此外，所获分区通常仅以解剖轴为参照进行标注，且未对其功能特征进行详细阐释，这一局限制约了对大脑功能的深入解析。为破解上述难题，本文提出一种带功能特征解析的深度聚类分区（Deep Clustering Parcellation with Functional Characterization, DCPFC）框架，并将其应用于两个公开开放的静息态fMRI数据集的个体及群体纹状体亚区分析。该框架引入深度聚类网络以提取并重构非线性特征，同时通过迭代优化同步完成聚类任务。此外，本文还提出一种融合流形距离的空间约束机制，以保障所得分区的空间连续性。尤为关键的是，通过将重构后的聚类质心映射至大脑皮层，可进一步解码这些纹状体分区的功能特征。与其他主流的基于连接组的分区（connectivity-based parcellation, CBP）方法相比，本文所提框架在可复现性与空间连续性上表现更优，且功能同质性与之相当。进一步地，纹状体被划分为6个分区，其划分依据涵盖躯体运动网络、注意网络（背侧与腹侧）及额顶网络。研究发现，认知灵活性与延迟折扣行为与这些分区的体积存在显著相关性，这印证了本框架能够有效捕捉大脑与认知相关的功能拓扑结构。本框架生成的具备功能可解释性的纹状体亚区，不仅进一步深化了我们对大脑的认知，更为临床研究的推进提供了坚实基础。