Search

Background and Objective: Leukocyte classification and cytometry have wide applications in medical domain, previous researches usually exploit machine learning techniques to classify leukocytes automatically. However, constrained by the past development of machine learning techniques, for example, extracting distinctive features from raw microscopic images are difficult, the widely used SVM classifier only has relative few parameters to tune, these methods cannot efficiently handle fine-grained classification cases when the white blood cells have up to 40 categories., Methods: Based on deep learning theory, a systematic study is conducted on finer leukocyte classification in this paper. A deep residual neural network based leukocyte classifier is constructed at first, which can imitate the domain expert's cell recognition process, and extract salient features robustly and automatically. Then the deep neural network classifier's topology is adjusted according to the prior knowledge of white blood cell test. After that the microscopic image dataset with almost one hundred thousand labeled leukocytes belonging to 40 categories is built, and combined training strategies are adopted to make the designed classifier has good generalization ability., Results: The proposed deep residual neural network based classifier was tested on microscopic image dataset with 40 leukocyte categories. It achieves top-1 accuracy of 77.80%, top-5 accuracy of 98.75% during the training procedure. The average accuracy on the test set is nearly 76.84%., Conclusions: This paper presents a fine-grained leukocyte classification method for microscopic images, based on deep residual learning theory and medical domain knowledge. Experimental results validate the feasibility and effectiveness of our approach. Extended experiments support that the fine-grained leukocyte classifier could be used in real medical applications, assist doctors in diagnosing diseases, reduce human power significantly., (Copyright © 2018 Elsevier B.V. All rights reserved.)

Data augmentation has been proved effective in training deep models. Existing data augmentation methods tackle fine-grained problem by blending image pairs and fusing corresponding labels according to the statistics of mixed pixels, which produces additional noise harmful to the performance of networks. Motivated by this, we present a simple yet effective cross ensemble knowledge distillation (CEKD) model for fine-grained feature learning. We innovatively propose a cross distillation module to provide additional supervision to alleviate the noise problem, and propose a collaborative ensemble module to overcome the target conflict problem. The proposed model can be trained in an end-to-end manner, and only requires image-level label supervision. Extensive experiments on widely used fine-grained benchmarks demonstrate the effectiveness of our proposed model. Specifically, with the backbone of ResNet-101, CEKD obtains the accuracy of 89.59%, 95.96% and 94.56% in three datasets respectively, outperforming state-of-the-art API-Net by 0.99%, 1.06% and 1.16%. [ABSTRACT FROM AUTHOR]

We present a novel active learning approach for shape cosegmentation based on graph convolutional networks (GCNs). The premise of our approach is to represent the collections of three-dimensional shapes as graph-structured data, where each node in the graph corresponds to a primitive patch of an oversegmented shape, and is associated with a representation initialized by extracting features. Then, the GCN operates directly on the graph to update the representation of each node based on a layer-wise propagation rule, which aggregates information from its neighbors, and predicts the labels for unlabeled nodes. Additionally, we further suggest an active learning strategy that queries the most informative samples to extend the initial training samples of GCN to generate more accurate predictions of our method. Our experimental results on the Shape COSEG dataset demonstrate the effectiveness of our approach. [ABSTRACT FROM AUTHOR]

Due to the inefficiency of gradient-based iterative ways in network training, randomization-based neural networks usually offer non-iterative closed form solutions. The random vector functional link (RVFL) and extreme learning machine (ELM) are two popular randomized networks which provide us unified frameworks for both regression and multi-class classification. Currently, existing studies on RVFL and ELM focused mainly on supervised tasks even though we usually have only a small number of labeled samples but a large number of unlabeled samples. Therefore, it is necessary to make both models appropriately utilize both labeled and unlabeled samples; that is, we should develop their semi-supervised extensions. In this paper, we propose a joint optimization framework to semi-supervised RVFL and ELM networks. In the formulated JOSRVFL (jointly optimized semi-supervised RVFL) and JOSELM, the output weight matrix and the label indicator matrix of the unlabeled samples can be jointly optimized in an iterative manner. We provide a novel approach to optimize the JOSRVFL and JOSELM objective functions. Extensive experiments on benchmark data sets and Electroencephalography-based emotion recognition tasks showed the excellent performance of the proposed JOSRVFL and JOSELM models. Moreover, because the direct input–output connections help to regularize the randomization, JOSRVFL could obtain superior performance to JOSELM in most cases. • A joint optimization framework to semi-supervised RVFL and ELM is proposed. • In JOSRVFL and JOSELM, we jointly estimate model variables and unlabeled samples. • We evaluate JOSRVFL and JOSELM on both benchmark and emotional EEG data sets. • With the input-output links, JOSRVFL often obtains superior performance to JOSELM. [ABSTRACT FROM AUTHOR]

Subspace learning represents a group of algorithms to project high dimensional data onto a low dimensional subspace in order to retain the desirable properties and simultaneously reduce the dimensionality. In graph-based subspace learning algorithms, the quality of graph affects the performance of projection matrix learning a lot. Especially when data is noisy or even grossly corrupted, we cannot guarantee that the constructed graph can accurately depict the inner structure of data. Additionally, the widely used two-stage paradigm of learning the projection matrix on a given graph isolates the connection between these two stages. In this paper, we propose a general framework to get rid of these disadvantages by the inspiration of low-rank matrix recovery and spectral regression-based subspace learning. Concretely, by jointly optimizing the objectives of low-rank representation and spectral regression, on one hand we can automatically get the recovered data from noise and the graph affinity matrix and on the other hand we can perform the projection matrix learning. The formulated joint low-rank and subspace learning (JLRSL) framework can be efficiently optimized by the augmented Lagrange multiplier method. We evaluate the performance of JLRSL by conducting extensive experiments on representative benchmark data sets and the results show that the low-rank learning can greatly facilitate the process of subspace learning, leading to robust feature extraction. Moreover, comparison with the state-of-the-arts is performed. [ABSTRACT FROM AUTHOR]