Michael Barnathan

This presentation was given based on the work that we did classifying Parkinson's Disease symptoms at BioMotion Suite.

Sole author. Barnathan, M. Mammographic Segmentation Using WaveCluster. Algorithms. 2012; 5(3):318-329.

I've become somewhat of an expert on the WaveCluster algorithm, since I authored the only publicly accessible implementation of it (as of mid-2012) and made a number of extensions to the basic algorithm which greatly increase its power.

Published in Neuroimage. 2011 Sep 15;58(2):537-48

My doctoral dissertation introduces TWave, the first comprehensive framework for high-order classification, clustering, compression, feature extraction, and latent concept discovery. Pioneering the use of tensor factorizations in biomedical image analysis, the principal application was in mining motor task fMRI data. Key insights include the formulation of grid clustering (on a grid of arbitrary dimensionality and order) as a locality-sensitive instance of latent semantic analysis, the fusion… Show more

My doctoral dissertation introduces TWave, the first comprehensive framework for high-order classification, clustering, compression, feature extraction, and latent concept discovery. Pioneering the use of tensor factorizations in biomedical image analysis, the principal application was in mining motor task fMRI data. Key insights include the formulation of grid clustering (on a grid of arbitrary dimensionality and order) as a locality-sensitive instance of latent semantic analysis, the fusion of wavelets with tensor factorizations to achieve a methodology which is sensitive to both global patterns and local neighborhoods at multiple resolutions, and the extension of WaveCluster into a versatile high-order, global, real-valued image clustering algorithm called TWaveCluster dominated by WaveCluster's linear runtime. Additional results include a novel tensor-theoretic formulation of the multidimensional separable wavelet transform which permits application to datasets of arbitrary order, a margin-sensitive variant of WaveCluster using Lloyd iteration to deform the grid cells prior to analysis (eliminating the potentially massive partial volume effect inherent in the naive WaveCluster algorithm), and the development of novel image analyses, including a texture-based computer-assisted diagnosis system achieving up to 89% accuracy in classification of brain tumor MRIs. From a 9.2 GB 6th-order digital-opposition task fMRI dataset, TWave derived subject handedness as a latent concept without human intervention, was capable of automatically identifying and distinguishing spatiotemporal activation patterns in the frontal and motor regions of the brain corresponding to planning to move and actual movement, unambiguously demonstrated the existence of ipsilateral motor activation in our dataset, and yet consumed two orders of magnitude less time and space vs. existing methods of image analysis. Show less

I am the second author on this paper. Among my first publications in the biomedical domain, this paper presents a methodology for topological analysis of the breast ductal network using Prufer and depth-first tree encodings with applications to diagnosis of breast cancer and other radiological abnormalities of the breast (with high discriminative value, as these conditions often distort the ductal architecture and thus radically alter the branching patterns, sometimes even before a gross lesion… Show more

I am the second author on this paper. Among my first publications in the biomedical domain, this paper presents a methodology for topological analysis of the breast ductal network using Prufer and depth-first tree encodings with applications to diagnosis of breast cancer and other radiological abnormalities of the breast (with high discriminative value, as these conditions often distort the ductal architecture and thus radically alter the branching patterns, sometimes even before a gross lesion is visible on mammography). The trees are manually segmented in this work; another paper I co-authored three years later would introduce an automatic segmentation method using a set of hybrid local features. Show less

In this work I fuse texture and topological analysis of branching patterns of the breast ductal network to match datasets of galactograms and mammograms. The methods used to assess branching were developed in my previous paper, but the novelty in this paper is nontrivial.

I'm particularly fond of the results of this paper, which utilizes keyblock texture learning (or "image patches", as they're now called) and vector quantization to distinguish between 11 types of normal and abnormal brain tissue in T1 and T2 MRIs of a transgenic mouse with a large medulloblastoma, including vascularization, peritumoral edema, and tissue necrosis - these imaging features are important as they can be used to grade human gliomas.

(I generally have some objections to using… Show more

I'm particularly fond of the results of this paper, which utilizes keyblock texture learning (or "image patches", as they're now called) and vector quantization to distinguish between 11 types of normal and abnormal brain tissue in T1 and T2 MRIs of a transgenic mouse with a large medulloblastoma, including vascularization, peritumoral edema, and tissue necrosis - these imaging features are important as they can be used to grade human gliomas.

(I generally have some objections to using transgenic animal tumor models in imaging studies when we're unfortunate enough to possess large volumes of human tumor imaging data, but this aspect of the work was out of my hands)

I had no idea how close I was working to the ideas that would lead to vision transformers here. Show less