A Research manuscript
The research report <K-parcel: Spatially-Continuous Brain Parcellation> could be downloaded here.
K-parcel Spatially-Continous Brain Parcellation.pdf
This is an unpublished manuscript and is considered as an exercise of research skill learning for my master's research supervised by Prof. Martin Ester and Prof. Maxwell Libbrecht.
A meta comment:
the proposed approach is somehow similar to the work of ReNA
Recursive nearest agglomeration (ReNA): fast clustering for approximation of structured signals
This paper shares the same intuiton but is much better than my work, in many aspects, and is rigorously evaluated. And it is fast.
Hoyos-Idrobo et al, 2019, IEEE transactions of on Pattern Analysis and Machine Intelligence
Functional parcellation purposes to divide a spatial activity map, such as brain images, into homogeneous and functionally coherent regions, which serve with fundamental importance as applications to analysis and interpretation.
Previous practice on this task choose to do clustering in the feature space while ignoring the spatial proximity, and consider features by calculating simple statistical measure such as correlation. We argue that those widely adopted clustering methods often results in not spatially continous and unsatisfactory parcellations in inevitable noisy settings in brain images, questioning the reliability of parcellation.
what this research is about:
Traditional anatomy-based brain parcellations often fail to capture he intrinsic functional coherence and thus comes the rise of data-driven parcellation methods based on clustering algorithms.
However, Current clustering algorithm is not good enough and in terms of the main evaluation metrics, algorithms generally give contrast results in different metrics: a algorithm performs well in one metric and give poorest performance in another metric. There is no algorithm that could satisfy all these metrics at the same time.
And One very important issue is that, existing method are especially fragile to noise which is inherent in current brain imaging technology.
We propose K-parcel, a simple clustering algorithm which could be viewed as a constrained K-means that guarantee to obtain spatially-connected and highly coherent functional parcellations on a spatial activity map. K-parcel is able to group voxels with coherent feature patterns and robust to noisy signals which leads to the aforementioned problems in practice.
The proposed method is compared with several alternative approaches through simulation studies, and is applied to real brain images. Experimental results demonstrate the superiority and robustness of K-parcel algorithm that it could find spatially-connected and homogeneous parcellation even in high noisy settings.