Sayan D. Pathak

Regression forests for efficient anatomy detection and localization in computed tomography scans

This paper proposes a new algorithm for the efficient, automatic detection and localization of multiple anatomical structures within three-dimensional computed tomography (CT) scans. Applications include selective retrieval of patients images from PACS systems, semantic visual navigation and tracking radiation dose over time. The main contribution of this work is a new, continuous parametrization of the anatomy localization problem, which allows it to be addressed effectively by multi-class random regression forests. Regression forests are similar to the more popular classification forests, but trained to predict continuous, multi-variate outputs, where the training focuses on maximizing the confidence of output predictions. A single pass of our probabilistic algorithm enables the direct mapping from voxels to organ location and size. Quantitative validation is performed on a database of 400 highly variable CT scans. We show that the proposed method is more accurate and robust than techniques based on efficient multi-atlas registration and template-based nearest-neighbor detection. Due to the simplicity of the regressors context-rich visual features and the algorithms parallelism, these results are achieved in typical run-times of only ∼4 s on a conventional single-core machine.

Clustering of spatial gene expression patterns in the mouse brain and comparison with classical neuroanatomy

Spatial gene expression profiles provide a novel means of exploring the structural organization of the brain. Computational analysis of these patterns is made possible by genome-scale mapping of the C57BL/6J mouse brain in the Allen Brain Atlas. Here we describe methodology used to explore the spatial structure of gene expression patterns across a set of 3041 genes chosen on the basis of consistency across experimental observations (N=2). The analysis was performed on smoothed, co-registered 3D expression volumes for each gene obtained by aggregating cellular resolution image data. Following dimensionality and noise reduction, voxels were clustered according to similarity of expression across the gene set. We illustrate the resulting parcellations of the mouse brain for different numbers of clusters (K) and quantitatively compare these parcellations with a classically-defined anatomical reference atlas at different levels of granularity, revealing a high degree of correspondence. These observations suggest that spatial localization of gene expression offers substantial promise in connecting knowledge at the molecular level with higher-level information about brain organization.

A discriminative-generative model for detecting intravenous contrast in CT images

This paper presents an algorithm for the automatic detection of intravenous contrast in CT scans. This is useful e.g. for quality control, given the unreliability of the existing DICOM contrast metadata. The algorithm is based on a hybrid discriminative-generative probabilistic model. A discriminative detector localizes enhancing regions of interest in the scan. Then a generative classifier optimally fuses evidence gathered from those regions into an efficient, probabilistic prediction. The main contribution is in the generative part. It assigns optimal weights to the detected organs based on their learned degree of enhancement under contrast material. The model is robust with respect to missing organs, patients geometry, pathology and settings. Validation is performed on a database of 400 highly variable patients CT scans. Results indicate detection accuracy greater than 91% at approximately 1 second per scan.

Can I foresee the success of my meetup group

Success of Meetup groups is of utmost importance for the members who organize them. Given a wide variety of such groups, a single metric may not be indicative of success for different groups; rather, success measure should be specific to the interest of a group. In this paper, accounting for the group diversity, we systematically define Meetup group success metrics and use them to generate labels for our machine learnt models. We crawl the Meetup dataset for three US cities namely New York, Chicago and San Francisco over a period of 8 months. The data study reveals the key players (such as core members, new members etc.) behind the success of the Meetup groups. This study leverages semantic, syntactic, temporal and location based features to discriminate between successful and unsuccessful groups. Finally, we present a model to predict success of the Meetup groups with high accuracy (0.81 with AUC = 0.86). Our approach generalizes well across groups, categories and cities. Additionally, the model performs reasonably well for new groups with little history (cold start problem), exhibiting high accuracy for the cross city validation.

Complementary Usage of Tips and Reviews for Location Recommendation in Yelp

Location-based social networks (LBSNs) allow users to share the locations that they have visited with others in a number of ways. LBSNs like Foursquare allow users to ‘check in’ to a location to share their locations with their friends. However, in Yelp, users can engage with the LBSN via modes other than check-ins. Specifically, Yelp allows users to write ‘tips’ and ‘reviews’ for the locations that they have visited. The geo-social correlations in LBSNs have been exploited to build systems that can recommend new locations to users. Traditionally, recommendation systems for LBSNs have leveraged check-ins to generate location recommendations. We demonstrate the impact of two new modalities - tips and reviews, on location recommendation. We propose a graph based recommendation framework which reconciles the ‘tip’ and ‘review’ space in Yelp in a complementary fashion. In the process, we define novel intra-user and intra-location links leveraging tip and review information, leading to a 15% increase in precision over the existing approaches.

Scalable Deep Document / Sequence Reasoning with Cognitive Toolkit

Deep Neural Networks (DNNs) have revolutionized the way that machines understand language and have allowed us to create models that answer textual questions, translate pairs of languages, and intelligently compare document corpora. At the heart of these successes lie core techniques that fall into the area of sequence understanding. While powerful, dealing with variable-size sequences in DNNs requires deep understanding and experience in creating such networks, which can be daunting to many scientists and engineers. This tutorial will focus on introducing core concepts, end-to-end recipes, and key innovations facilitated by the cross-platform fully open-source Cognitive Toolkit (formerly called CNTK) with superior scalability (up to 1000 GPUs) for very large data corpora. Specifically, we will present tutorials on basic sequence understanding, intermediate sequence-to-sequence translation (both with and without attention), and the advanced Reasoning Network (ReasoNet) which has achieved industry-leading results in reading comprehension.

Extracting Entities of Interest from Comparative Product Reviews

This paper presents a deep learning based approach to extract product comparison information out of user reviews on various e-commerce websites. Any comparative product review has three major entities of information: the names of the products being compared, the user opinion (predicate) and the feature or aspect under comparison. All these informing entities are dependent on each other and bound by the rules of the language, in the review. We observe that their inter-dependencies can be captured well using LSTMs. We evaluate our system on existing manually labeled datasets and observe out-performance over the existing Semantic Role Labeling (SRL) framework popular for this task.

Gene expression based mouse brain parcellation using Markov random field regularized non-negative matrix factorization

Understanding the geography of genetic expression in the mouse brain has opened previously unexplored avenues in neuroinformatics. The Allen Brain Atlas (www.brain-map.org) (ABA) provides genome-wide colorimetric in situ hybridization (ISH) gene expression images at high spatial resolution, all mapped to a common three-dimensional 200μm3 spatial framework defined by the Allen Reference Atlas (ARA) and is a unique data set for studying expression based structural and functional organization of the brain. The goal of this study was to facilitate an unbiased data-driven structural partitioning of the major structures in the mouse brain. We have developed an algorithm that uses nonnegative matrix factorization (NMF) to perform parts based analysis of ISH gene expression images. The standard NMF approach and its variants are limited in their ability to flexibly integrate prior knowledge, in the context of spatial data. In this paper, we introduce spatial connectivity as an additional regularization in NMF decomposition via the use of Markov Random Fields (mNMF). The mNMF algorithm alternates neighborhood updates with iterations of the standard NMF algorithm to exploit spatial correlations in the data. We present the algorithm and show the sub-divisions of hippocampus and somatosensory-cortex obtained via this approach. The results are compared with established neuroanatomic knowledge. We also highlight novel gene expression based sub divisions of the hippocampus identified by using the mNMF algorithm.