Vincent C. Gaudet

Manifold learning is used for dimensionality reduction, with the goal of finding a projection subspace to increase and decrease the inter- and intraclass variances, respectively. However, a bottleneck for subspace learning methods often arises from the high dimensionality of datasets. In this paper, a hierarchical approach is proposed to scale subspace learning methods, with the goal of improving classification in large datasets by a range of 3% to 10%. Different combinations of methods are studied. We assess the proposed method on five publicly available large datasets, for different eigen-value based subspace learning methods such as linear discriminant analysis, principal component analysis, generalized discriminant analysis, and reconstruction independent component analysis. To further examine the effect of the proposed method on various classification methods, we fed the generated result to linear discriminant analysis, quadratic linear analysis, k-nearest neighbor, and random forest classifiers. The resulting classification accuracies are compared to show the effectiveness of the hierarchical approach, reporting results of an average of 5% increase in classification accuracy.

It is common for a time series dataset to have missing values, and it is necessary to fill these missing elements before fitting any model for forecasting or prediction. Time series imputation remains a challenging task due to the existence of non-linear dependencies between current and past values. Conventional methods, such as deletion of rows containing missing values or filling them with the last observed value, add bias to the data and are therefore inefficient. There are situations where data is missing at consecutive points or random points in the dataset, and one particular method may not work well for all cases. In this paper, nine commonly used models in the field of imputation, based on tools of statistics, machine learning, and deep learning, are compared. Results show that Linear Memory Vector Gated Recurrent Unit (LIME-GRU) outperforms the other tested models by having the least Mean Square Error (MSE) and Root Mean Squared Error (RMSE). A predictive model to gauge the impact of imputation on prediction is also used to validate the findings. The results of the prediction model illustrate that with LIME-GRU, there was a 39% improvement in Average Aggregated Measure (AAGM) when compared with mode imputation on a particular test case.

Purpose: To evaluate random forests (RFs) as a supervised machine learning algorithm to classify amyloid brain PET as positive or negative for amyloid deposition and identify key regions of interest for stratification. Methods: The data set included 57 baseline ¹⁸F-florbetapir (Amyvid; Lilly, Indianapolis, IN) brain PET scans in participants with severe white matter disease, presenting with either transient ischemic attack/lacunar stroke or mild cognitive impairment from early Alzheimer disease, enrolled in a multicenter prospective observational trial. Scans were processed using the MINC toolkit to generate SUV ratios, normalized to cerebellar gray matter, and clinically read by 2 nuclear medicine physicians with interpretation based on consensus (35 negative, 22 positive). SUV ratio data and clinical reads were used for supervised training of an RF classifier programmed in MATLAB. Results: A 10,000-tree RF, each tree using 15 randomly selected cases and 20 randomly selected features (SUV ratio per region of interest), with 37 cases for training and 20 cases for testing, had sensitivity = 86% (95% confidence interval [CI], 42%–100%), specificity = 92% (CI, 64%–100%), and classification accuracy = 90% (CI, 68%–99%). The most common features at the root node (key regions for stratification) were (1) left posterior cingulate (1039 trees), (2) left middle frontal gyrus (1038 trees), (3) left precuneus (857 trees), (4) right anterior cingulate gyrus (655 trees), and (5) right posterior cingulate (588 trees). Conclusions: Random forests can classify brain PET as positive or negative for amyloid deposition and suggest key clinically relevant, regional features for classification.

Machine learning (ML) is increasingly used in medical imaging. This paper provides pilot data of decision trees and random forests (RFs) to predict if a ¹⁸ F-florbetapir brain positron emission tomography (PET) is positive or negative for amyloid deposition based on quantitative data analysis. The dataset included 55 ¹⁸ F-florbetapir brain PETs in participants with severe white matter disease and mild cognitive impairment (MCI), early Alzheimer's disease (AD) or transient ischemic events. The Montreal Cognitive Assessment (MoCA) score was known for each participant. All PET images were processed using the MINC toolkit to extract standardized uptake value ratios (SUVRs) for 59 regions of interest (features). Each PET was clinically read by 2 dual certified radiology/nuclear medicine physicians with final interpretation based on consensus. An initial study of RFs using conventional binary decision trees and PET quantitation suggests this is a powerful algorithm for PET classification as positive or negative for amyloid deposition. Preliminary data did not show improved results when a ternary RF approach was used. Finally, a soft-decision approach may be helpful to predict the MoCA score.

This article, the first in a two-part series, provides an introduction to machine learning in a nuclear medicine context. Part I in this series addresses the history of machine learning and describes common algorithms with illustrations of when these algorithms can be helpful in nuclear medicine. Part II focuses on current contributions of machine learning to our field and addresses future expectations, limitations as well as providing a critical appraisal of what machine learning can and cannot do.

Limited battery capacity and long charging time resulting in what is known as range anxiety have been major obstacles to the widespread adoption of electric vehicles. In addition to running out of battery power, some drivers are also concerned about the amount of time required to recharge their batteries (i.e., time anxiety). This paper focuses on these problems, proposing a Multi-Module Range Anxiety Reduction Scheme. The proposed scheme takes into account traffic density on roadways to provide an accurate computation of energy consumption to charging stations in order to overcome the driver's concern of being stranded en-route. Furthermore, it addresses the driver's concern about completing the recharging process in either minimum energy or time. Simulations are conducted to test and validate the proposed scheme.

The recent trend toward increasingly deep convolutional neural networks (CNNs) leads to a higher demand of computational power and memory storage. Consequently, the deployment of CNNs in hardware has become more challenging. In this paper, we propose an Intra-Kernel Regular (IKR) pruning scheme to reduce the size and computational complexity of the CNNs by removing redundant weights at a fine-grained level. Unlike other pruning methods such as Fine-Grained pruning, IKR pruning maintains regular kernel structures that are exploitable in a hardware accelerator. Experimental results demonstrate up to 10× parameter reduction and 7× computational reduction at a cost of less than 1% degradation in accuracy versus the unpruned case.