Hasan Ertan Cetingul

Sequential Hierarchical Agglomerative Clustering of White Matter Fiber Pathways

Objective: We consider the problem of clustering white matter fiber pathways, extracted from diffusion MRI data via tractography, into bundles that are consistent with the neuroanatomy. Methods: We cast this problem as clustering streams of data, and use a sequential framework to process one fiber at a time. Our method, named as sequential hierarchical agglomerative clustering (HAC), represents the clusters with parametric models, performs HAC of relatively small number of fibers only when the parameters need to be initialized and/or updated, and assigns the labels to the following streams of data according to the current models. Results: Experiments on phantom data evaluate the sensitivity of our method to initialization and parameter tuning, and show its advantages over alternative techniques. Experiments on real data demonstrate its efficacy and speed in clustering white matter fiber pathways into anatomically distinct bundles. Conclusion: Sequential HAC is a fast method that benefits from having a predefined number of clusters, and rapidly assigns labels to incoming data with high accuracy. It can be thought of as a mechanism that does clustering, while simultaneously accepting newly computed fibers; thereby, alleviating the burden of computing the distances between every pair of fibers in a tractogram. Significance: Sequential HAC is a practical tool that can interactively cluster fiber pathways and can be integrated into fiber tracking, which will be very useful for clinical researchers and neuroanatomists.

Learning Global and Cluster-Specific Classifiers for Robust Brain Extraction in MR Data

We present a learning-based framework for automatic brain extraction in MR images. It accepts single or multi-contrast brain MR data, builds global binary random forests classifiers at multiple resolution levels, hierarchically performs voxelwise classifications for a test subject, and refines the brain surface using a narrow-band level set technique on the classification map. We further develop a data-driven schema to improve the model performance, which clusters patches of co-registered training images and learns cluster-specific classifiers. We validate our framework via experiments on single and multi-contrast datasets acquired using scanners with different magnetic field strengths. Compared to the state-of-the-art methods, it yields the best performance with statistically significant improvement of the cluster-specific method (with a Dice coefficient of 97.6 ± 0.4 % and an average surface distance of 0.8 ± 0.1 mm) over the global method.