Jean-Baptiste Fiot

Efficient brain lesion segmentation using multi‐modality tissue‐based feature selection and support vector machines

Support vector machines (SVM) are machine learning techniques that have been used for segmentation and classification of medical images, including segmentation of white matter hyper-intensities (WMH). Current approaches using SVM for WMH segmentation extract features from the brain and classify these followed by complex post-processing steps to remove false positives. The method presented in this paper combines advanced pre-processing, tissue-based feature selection and SVM classification to obtain efficient and accurate WMH segmentation. Features from 125 patients, generated from up to four MR modalities [T1-w, T2-w, proton-density and fluid attenuated inversion recovery(FLAIR)], differing neighbourhood sizes and the use of multi-scale features were compared. We found that although using all four modalities gave the best overall classification (average Dice scores of 0.54  ±  0.12, 0.72  ±  0.06 and 0.82  ±  0.06 respectively for small, moderate and severe lesion loads); this was not significantly different (p = 0.50) from using just T1-w and FLAIR sequences (Dice scores of 0.52  ±  0.13, 0.71  ±  0.08 and 0.81  ±  0.07). Furthermore, there was a negligible difference between using 5 × 5 × 5 and 3 × 3 × 3 features (p = 0.93). Finally, we show that careful consideration of features and pre-processing techniques not only saves storage space and computation time but also leads to more efficient classification, which outperforms the one based on all features with post-processing.

Electricity Demand Forecasting by Multi-Task Learning

We explore the application of kernel-based multi-task learning techniques to forecast the demand of electricity measured on multiple lines of a distribution network. We show that recently developed output kernel learning techniques are particularly well suited to solve this problem, as they allow to flexibly model the complex seasonal effects that characterize electricity demand data, while learning and exploiting correlations between multiple demand profiles. We also demonstrate that kernels with a multiplicative structure yield superior predictive performance with respect to the widely adopted (generalized) additive models. This paper is based on residential and industrial smart meter data provided by the Irish Commission for Energy Regulation.

Local vs global descriptors of hippocampus shape evolution for alzheimer’s longitudinal population analysis

In the context of Alzheimers disease (AD), state-of-the-art methods separating normal control (NC) from AD patients or CN from progressive MCI (mild cognitive impairment patients converting to AD) achieve decent classification rates. However, they all perform poorly at separating stable MCI (MCI patients not converting to AD) and progressive MCI. Instead of using features extracted from a single temporal point, we address this problem using descriptors of the hippocampus evolutions between two time points. To encode the transformation, we use the framework of large deformations by diffeomorphisms that provides geodesic evolutions. To perform statistics on those local features in a common coordinate system, we introduce an extension of the Karcher mean algorithm that defines the template modulo rigid registrations, and an initialization criterion that provides a final template leading to better matching with the patients. Finally, as local descriptors transported to this template do not directly perform as well as global descriptors (e.g. volume difference), we propose a novel strategy combining the use of initial momentum from geodesic shooting, extended Karcher algorithm, density transport and integration on a hippocampus subregion, which is able to outperform global descriptors.

Enabling coupled models to predict the business impact of weather on electric utilities

Efficient, resilient, and safe operation of an electric utility is dependent on the local weather conditions at the scale of its infrastructure. This sensitivity to weather includes such factors as damage to distribution or transmission systems due to relative extremes in precipitation or wind, determining electricity demand and load, and power generation from renewable facilities. Hence, the availability of highly focused weather predictions has the potential to enable proactive planning for the effect of weather on utility systems. Often, such information is simply unavailable. The initial step to address this gap is the application of state-of-the-art physical weather models at the spatial scale of the utilitys infrastructure, calibrated to avoid this mismatch in predictability. The results of such a model are then coupled to a data-driven stochastic model to represent the weather impacts. The deployment of such methods requires an abstraction of the weather forecasting component to drive the model coupling.

A scalable demand and renewable energy forecasting system for distribution grids

Managing a reliable, renewable, and affordable power grid is a challenging task because the mix of power generating and consuming devices connected to the network continues to change. Improved forecasts help network operators respond to these changes and make data-driven decisions regarding, e.g., demand response and market operations. A system producing short-term energy forecasts of demand and renewable generation at multiple aggregation levels across the service territory of a distribution utility is presented. The system automates the process of ingesting and curating large amounts of data from multiple sources, such as high-resolution weather forecasts, SCADA (supervisory control and data acquisition) data and, smart meter data. This results in a richer and higher-quality data set which improves accuracy for residual demand forecasts because it enables the use of real-time data and the creation of detailed models for solar energy generation. Results of an operational deployment of the system on the service territory covered by the largest electric distribution utility in Vermont, Green Mountain Power, are presented.

Electricity demand forecasting by multi-task learning

We explore the application of kernel-based multi-task learning techniques to forecast the demand of electricity measured on multiple lines of a distribution network. We show that recently developed output kernel learning techniques are particularly well suited to solve this problem, as they allow to flexibly model the complex seasonal effects that characterize electricity demand data, while learning and exploiting correlations between multiple demand profiles. We also demonstrate that kernels with a multiplicative structure yield superior predictive performance with respect to the widely adopted (generalized) additive models. Our study is based on residential and industrial smart meter data provided by the Irish Commission for Energy Regulation (CER).

Structured Dimensionality Reduction for Additive Model Regression

Additive models are regression methods which model the response variable as the sum of univariate transfer functions of the input variables. Key benefits of additive models are their accuracy and interpretability on many real-world tasks. Additive models are however not adapted to problems involving a large number (e.g., hundreds) of input variables, as they are prone to overfitting in addition to losing interpretability. In this paper, we introduce a novel framework for applying additive models to a large number of input variables. The key idea is to reduce the task dimensionality by deriving a small number of new covariates obtained by linear combinations of the inputs, where the linear weights are estimated with regard to the regression problem at hand. The weights are moreover constrained to prevent overfitting and facilitate the interpretation of the derived covariates. We establish identifiability of the proposed model under mild assumptions and present an efficient approximate learning algorithm. Experiments on synthetic and real-world data demonstrate that our approach compares favorably to baseline methods in terms of accuracy, while resulting in models of lower complexity and yielding practical insights into high-dimensional real-world regression tasks. Our framework broadens the applicability of additive models to high-dimensional problems while maintaining their interpretability and potential to provide practical insights.

Combining imaging and clinical data in manifold learning: Distance-based and graph-based extensions of Laplacian Eigenmaps

Manifold learning techniques have been widely used to produce low-dimensional representations of patient brain magnetic resonance (MR) images. Diagnosis classifiers trained on these coordinates attempt to separate healthy, mild cognitive impairment and Alzheimers disease patients. The performance of such classifiers can be improved by incorporating clinical data available in most large-scale clinical studies. However, the standard non-linear dimensionality reduction algorithms cannot be applied directly to imaging and clinical data. In this paper, we introduce a novel extension of Laplacian Eigenmaps that allow the computation of manifolds while combining imaging and clinical data. This method is a distance-based extension that suits better continuous clinical variables than the existing graph-based extension, which is suitable for clinical variables in finite discrete spaces. These methods were evaluated in terms of classification accuracy using 288 MR images and clinical data (ApoE genotypes, Aβ42 concentrations and mini-mental state exam (MMSE) cognitive scores) of patients enrolled in the Alzheimers disease neuroimaging initiative (ADNI) study.