Semi-supervised Mode Research Articles

Weakly Supervised Semantic Segmentation via Box-driven Masking and Filling Rate Shifting.

Semantic segmentation has achieved huge progress via adopting deep Fully Convolutional Networks (FCN). However, the performance of FCN-based models severely rely on the amounts of pixel-level annotations which are expensive and time-consuming. Considering that bounding boxes also contain abundant semantic and objective information, an intuitive solution is to learn the segmentation with weak supervisions from the bounding boxes. How to make full use of the class-level and region-level supervisions from bounding boxes to estimate the uncertain regions is the critical challenge for the weakly supervised learning task. In this paper, we propose a mixture model to address this problem. First, we introduce a box-driven class-wise masking model (BCM) to remove irrelevant regions of each class. Moreover, based on the pixel-level segment proposal generated from the bounding box supervision, we calculate the mean filling rates of each class to serve as an important prior cue to guide the model ignoring the wrongly labeled pixels in proposals. To realize the more fine-grained supervision at instance-level, we further propose the anchor-based filling rate shifting module. Unlike previous methods that directly train models with the generated noisy proposals, our method can adjust the model learning dynamically with the adaptive segmentation loss. Thus it can help reduce the negative impacts from wrongly labeled proposals. Besides, based on the learned high-quality proposals with above pipeline, we explore to further boost the performance through two-stage learning. The proposed method is evaluated on the challenging PASCAL VOC 2012 benchmark and achieves 74.9% and 76.4% mean IoU accuracy under weakly and semi-supervised modes, respectively. Extensive experimental results show that the proposed method is effective and is on par with, or even better than current state-of-the-art methods. Code will be available at: https://github.com/developfeng/BCM.

A deep generative framework with embedded vector arithmetic and classifier for sample generation, label transfer, and clustering of single-cell data

Multiple-source single-cell datasets have accumulated quickly and need computational methods to integrate and decompose into meaningful components. Here, we present inClust (integrated clustering), a flexible deep generative framework that enables embedding auxiliary information, latent space vector arithmetic, and clustering. All functional parts are relatively modular, independent in implementation but interrelated at runtime, resulting in an all-in general framework that could work in supervised, semi-supervised, or unsupervised mode. We show that inClust is superior to most data integration methods in benchmark datasets. Then, we demonstrate the capability of inClust in the tasks of conditional out-of-distribution generation in supervised mode, label transfer in semi-supervised mode, and spatial domain identification in unsupervised mode. In these examples, inClust could accurately express the effect of each covariate, distinguish the query-specific cell types, or segment spatial domains. The results support that inClust is an excellent general framework for multiple-task harmonization and data decomposition.

Structure-Enriched Topology Learning For Cross-Domain Multi-Person Pose Estimation

Human pose estimation has been widely studied with much focus on supervised learning. However, in real applications, a pretrained pose estimation model usually needs be adapted to a novel domain without labels or with sparse labels. Existing domain adaptation methods cannot well deal with it since poses have flexible topological structures and need fine-grained local features. Aiming at the characteristics of human pose, we propose a novel domain adaptation method for multi-person pose estimation (MPPE) to alleviate the human-level shift. Firstly, the training samples of human poses are clustered into groups according to the posture similarity. Within the clustered space, we conduct three adaptation modules: Cross-Attentive Feature Alignment (CAFA), Intra-domain Structure Adaptation (ISA) and Adaptive Human-Topology Adaptation (AHTA). The CAFA adopts a bidirectional spatial attention mechanism to explore fine-grained local feature correlation between two humans, and thus to adaptively aggregate consistent features for adaptation. ISA only works in semi-supervised domain adaptation (SSDA) to exploit semantic relationship of corresponding keypoints for reducing the intra-domain bias. Importantly, we creatively propose an AHTA to enrich human topological knowledge for reducing the inter-domain discrepancy. Specifically, the pose structure and the cross-instance topological relations are modeled via graph networks. This flexible topology learning benefits the occluded or extreme pose inference. Extensive experiments are conducted on two popular benchmarks and additional two challenging datasets. Results demonstrate the competency of our method, which works in unsupervised or semi-supervised modes, compared with the existing supervised approaches.

A Comparative Study for Outlier Detection Methods in High Dimensional Text Data

Abstract Outlier detection aims to find a data sample that is significantly different from other data samples. Various outlier detection methods have been proposed and have been shown to be able to detect anomalies in many practical problems. However, in high dimensional data, conventional outlier detection methods often behave unexpectedly due to a phenomenon called the curse of dimensionality. In this paper, we compare and analyze outlier detection performance in various experimental settings, focusing on text data with dimensions typically in the tens of thousands. Experimental setups were simulated to compare the performance of outlier detection methods in unsupervised versus semi-supervised mode and uni-modal versus multi-modal data distributions. The performance of outlier detection methods based on dimension reduction is compared, and a discussion on using k-NN distance in high dimensional data is also provided. Analysis through experimental comparison in various environments can provide insights into the application of outlier detection methods in high dimensional data.

Machine learning based assessment of small-bodied fish tracking to evaluate spoiler baffle fish passage design

Fish passage research is important to mitigate the adverse effects of fragmented river habitats caused by waterway structures. The scale at which this research is undertaken varies from small-scale laboratory prototype studies to in-situ observations at various fish passage structures and bottlenecks. Using DeepLabCut, we introduce and evaluate a machine learning based workflow to track small-bodied fish in order to facilitate improved fish passage management. We specifically studied the behaviour and kinematics of Galaxias maculatus, a widespread diadromous Southern Hemisphere fish species. Upstream fish passage was studied in the presence of three different patches of spoiler baffles at an average water velocity of 0.4 m/s. In semi-supervised mode, the fish locations were extracted, and fish behaviour, such as swimming pathways and resting locations, was analysed based on extracted positions and recorded kinematic parameters. Individual fish behaviour and kinematic parameters were then used to assess the suitability of the three different spoiler baffle designs for enhancing fish passage. Using this technique, we were able to demonstrate where different spoiler baffle configurations resulted in significant differences in fish passage success and behaviour. For example, medium-spaced smaller baffles provided more accessible and uniform resting locations, which were required for efficient upstream passage. Results are discussed in relation to fish passage management at small instream structures.

Semi-supervised Mode Classification of Inter-city Trips from Cellular Network Data

Good knowledge of travel patterns is essential in transportation planning. Cellular network data as a large-scale passive data source provides billions of daily location updates allowing us to observe human mobility with all travel modes. However, many transport planning applications require an understanding of travel patterns separated by travel mode, requiring the classification of trips by travel mode. Most previous studies have used rule-based or geometric classification, which often fails when the routes for different modes are similar or supervised classification, requiring labelled training trips. Sufficient amounts of labelled training trips are unfortunately often unavailable in practice. We propose semi-supervised classification as a novel approach of classifying large sets of trips extracted from cellular network data in inter-city origin–destination pairs as either using road or rail. Our methods require no labelled trips which is an important advantage as labeled data is often not available in practice. We propose three methods which first label a small share of trips using geometric classification. We then use structures in a large set of unlabelled trips using a supervised classification method (geometric-labelling), iterative semi-supervised training (self-labelling) and by transferring information between origin–destination pairs (continuity-labelling). We apply the semi-supervised classification methods on a dataset of 9545 unlabelled trips in two inter-city origin–destination pairs. We find that the methods can identify structures in the cells used during trips in the unlabelled data corresponding to the available route alternatives. We validate the classification methods using a dataset of 255 manually labelled trips in the two origin–destination pairs. While geometric classification misclassifies 4.2% and 5.6% of the trips in the two origin–destination pairs, all trips can be classified correctly using semi-supervised classification.

Semi-supervised robust training with generalized perturbed neighborhood

Adversarial examples have been shown to be a severe threat to deep neural networks (DNNs). One of the most effective adversarial defense methods is adversarial training (AT) through minimizing the adversarial risk Radv, which encourages both the benign example x and its adversarially perturbed neighborhoods within the ℓp-ball to be predicted as the ground-truth label. In this paper, we propose a novel defense method, the robust training (RT), by jointly minimizing two separated risks (i.e., Rstand and Rrob), which are with respect to the benign example and its neighborhoods, respectively. The motivation is to explicitly and jointly enhance the accuracy and the adversarial robustness. We prove that Radv is upper-bounded by Rstand+Rrob, which implies that RT has similar effect as AT. Intuitively, minimizing the standard risk enforces the benign example to be correctly predicted, while the robust risk minimization encourages the predictions of the neighbor examples to be consistent with the prediction of the benign example. Besides, since Rrob is independent of the ground-truth label, RT is naturally extended to the semi-supervised mode (i.e., SRT), to further enhance its effectiveness. Moreover, we extend the ℓp-bounded neighborhood to a general case, which covers different types of perturbations, such as the pixel-wise (i.e., x+δ) or the spatial perturbation (i.e., Ax+b). Extensive experiments on benchmark datasets not only verify the superiority of the proposed SRT to state-of-the-art methods for defending pixel-wise or spatial perturbations separately but also demonstrate its robustness to both perturbations simultaneously. Our work may shed the light on the understanding of universal model robustness and the potential of unlabeled samples. The code for reproducing main results is available at https://github.com/THUYimingLi/Semi-supervised_Robust_Training.

Transfer learning aid the prediction of sintering densification

In powder metallurgy engineering, the master sintering curve (MSC) is crucial for estimating the mechanical properties of sintered products and optimizing sintering process parameters. A rapid evaluation method, that is, a domain-adversarial neural network, is established in the field of sintering to transfer learning from Al2O3 to SiC, concerning the effect of the heating rate on the sintering densification, in which one material (e.g. SiC) lacks an MSC and the material (e.g. Al2O3) has an overabundance of MSCs. In the unsupervised mode, we can roughly predict the densification evolution of SiC based on the MSC data of Al2O3. In the semi-supervised mode, prediction accuracy gradually increases with the increase of Al2O3 data and approaches a certain upper limit. Compared with the traditional sintering density prediction models, the proposed approach can provide an effective and rapid solution to the problem of data scarcity in the sintering field.

Heuristic models and methods for application of the kohonen neural network in the intellectual system of medical-sociological monitoring

For increasing the speed of making classification decisions, the possibilities of supervised deep neural networks training are used in systems of medical and sociological monitoring. But the data obtained during the monitoring process are semi-structured and -labelled. This greatly reduces speed and increases supervised machine-training error. The paper proposes heuristic models and methods of using the Kohonen network to increase the training rate on the semi-supervised mode without losing the accuracy level of data classification in intelligent systems of medical and sociological monitoring. More specifically, proposed tools include improved models for the presentation of semi-structured and -labelled medical and sociological monitoring data in the spaces of properties and features due to formalization of procedure of the detailed data aggregation, as well as the semi-supervised training of the Kohonen network using the method of matching existing pre-labelled data with the obtained labels of features clusters. In addition, the method of adjusting weight coefficients in the process of training the Kohonen neural network through the use of a modified genetic algorithm to classify medical and sociological monitoring data was further developed. The proposed modification of the genetic algorithm consists in using the method of matching labels of classes and clusters in assessment the adaptability of chromosomes in the population generated for each example from training sample of medical and sociological monitoring data. It is proposed for evaluating the adaptability of chromosomes to calculate the completeness of the classification as the ratio of the number of found representatives of the positive class to all representatives of the positive class. When performing the selection procedure on the basis of the obtained adaptability score, the two most adapted (“winner” and “vice-winner”) chromosomes are selected and crossed, whereas in contrast, the two least adapted(“loser” and “vice-loser”) chromosomes are deleted. The crossing is implemented as the exchange of genes between the winner and vice-winner chromosomes, taking into account a randomly selected crossing-over point. Upon mutation, a random number distributed according to a uniform law is added to each chromosome gene, which is an auxiliary method for creating a new chromosome to prevent degeneration of the population. The proposed heuristic models and methods are the basis for the implementation of a separate module as part of the intellectual system of medical and sociological monitoring. Testing the modified intellectual system using well-known test examples from the machine training database and real medical and sociological data showed an increase in the training speed of the Kohonen network with a given level of classification accuracy. Thus, the introduction of the proposed tools allows increasing the efficiency of classification decisions without losing their level of accuracy in intelligent systems of medical and sociological monitoring.

A Model-Based Unsupervised Deep Learning Method for Low-Dose CT Reconstruction

Low-dose CT (LDCT) is of great significance due to the concern about the potential radiation risk. With the fast development of deep learning, neural networks have become powerful tools in LDCT enhancement. Current deep neural networks for LDCT reconstruction are often trained with paired LDCT dataset and normal-dose CT (NDCT) dataset. However, high quality NDCT dataset paired with LDCT dataset is expensive to acquire or even not available sometimes in reality. In this work, we proposed an unsupervised model-based deep learning (MBDL) for LDCT reconstruction. The network is trained based on group-wise maximum a posterior (G-MAP) loss function with LDCT dataset only. The MBDL is a general framework. It also allows us to combine with supervised training if a small number of paired NDCT dataset accessible to help optimizing the network parameters, i.e. works in a semi-supervised mode. During inference, LDCT images are reconstructed end-to-end by the trained network. We verified the proposed method with simulated projection data from clinical CT images. The proposed method restrained noise well while restoring anatomical structures and it achieved better results than model-based iterative reconstruction (MBIR) with significantly less computational cost. The performances of MBDL were further enhanced by integrating a small paired NDCT dataset for semi-supervised training. The results suggested that MBDL is an efficient and flexible method for LDCT deep learning based reconstruction in the situations lacking of enough high quality NDCT data.

Semi-Supervised Learning-Based Image Denoising for Big Data

In this paper, the research of image noise reduction based on semi-supervised learning is carried out, and the neural network is used to reduce the noise of the image, so as to achieve more stable and good image display ability. Based on the convolutional neural network algorithm, the role of activation function optimization network is studied, combined with semi-supervised learning modes such as multi-feature extraction technology, to learn and extract the key features of the input image. Semi-supervised residual learning based on convolutional network is a good image denoising and denoising network model. Compared with other excellent denoising algorithms, it has very good results. At the same time, it greatly improves the image noise pollution and makes the image details clearer. At the same time, compared with other image denoising algorithms, this algorithm can show a good peak signal-to-noise ratio under various noise standard deviations. Through the research in this article, it is verified that the improved convolutional neural network denoising model and multi-feature extraction technology have strong advantages in image denoising.

Anomaly Detection and Diagnosis In Manufacturing Systems

Multivariate sensor data collected from manufacturing and process industries represents actual operational behavior and can be used for predictive maintenance of the plants. Anomaly detection and diagnosis, that forms an integral part of predictive maintenance, in industrial systems is however challenging due to their complex behavior, interactions among sensors, corrective actions of control systems and variability in anomalous behavior. While several statistical techniques for anomaly detection have been in use for a long time, these are not particularly suited for temporal (or contextual) anomalies that are characteristic of multivariate time series sensor data. On the other hand, several machine learning and deep learning techniques for anomaly detection gained significant interest in the recent years. Further, anomaly diagnosis that involves localization of the faults did not receive much attention. In this work, we compare the anomaly detection and diagnosis capabilities, in semi-supervised mode, of several statistical, machine learning and deep learning techniques on two systems viz. the interacting quadruple tank system and the continuous stirred tank reactor (CSTR) system both of which are representative of the complexity of large industrial systems. The techniques studied include principal component analysis (PCA), Mahalanobis distance (MD), one-class support vector machine (OCSVM), isolation forest, elliptic envelope, dense auto-encoder and long short term memory auto-encoder (LSTM AE). The study revealed that MD and LSTM-AE have the highest anomaly detection capability, followed closely by PCA and OCSVM. The above techniques also exhibited good diagnosis capability. The study indicates that statistical techniques in spite of their simplicity could be as powerful as machine learning and deep learning techniques, and may be considered for anomaly detection and diagnosis in manufacturing systems.

Semi-supervised Aspect-level Sentiment Classification Model based on Variational Autoencoder

Aspect-level sentiment classification aims to predict the sentiment of a text in different aspects and it is a fine-grained sentiment analysis task. Recent work exploits an Attention-based Long Short-Term Memory Network to perform aspect-level sentiment classification. Most previous work are based on supervised learning that needs a large number of labeled samples, but the problem is that only a limited subset of data samples are labeled in practical applications. To solve this problem, we propose a novel Semi-supervised Aspect Level Sentiment Classification Model based on Variational Autoencoder (AL-SSVAE) for semi-supervised learning in the aspect-level sentiment classification. The AL-SSVAE model inputs a given aspect to an encoder a decoder based on a variational autoencoder (VAE), and it also has an aspect level sentiment classifier. It enables the attention mechanism to deal with different parts of a text when different aspects are taken as input as previous methods. Due to that the sentiment polarity of a word is usually sensitive to the given aspect, a single vector for a word is problematic. Therefore, we propose the aspect-specific word embedding learning from a topical word embeddings model to express a word and also append the corresponding sentiment vector into the word input vector. We compare our AL-SSVAE model with several recent aspect-level sentiment classification models on the SemEval 2016 dataset. The experimental results indicate that the proposed model is able to capture more accurate semantics and sentiment for the given aspect and obtain better performance on the task of the aspect level sentiment classification. Moreover, the AL-SSVAE model is able to learn with the semi-supervised mode in the aspect level sentiment classification, which enables it to learn efficiently using less labeled data.

Investigating the influence of habitat structure and hydraulics on tropical macroinvertebrate communities

The influences of habitat structure and hydraulics on tropical macroinvertebrate communities were investigated in two foothill rivers of the Udzungwa Mountains (United Republic of Tanzania) to assist future Environmental Flow Assessments (EFAs). Macroinvertebrate samples, hydraulic variables and habitat structure were collected at the microhabitat scale (n=90). Macroinvertebrate communities were first delineated (i.e. clustered) through Poisson and negative binomial mixture models for count data in a semi-supervised mode by taking into account the sampled river. Then, genetically optimised Multi-Layer Perceptrons (MLPs) were used to identify the relationship of the most relevant variables with the delineated communities. Between the three delineated communities exclusively one community was shared between both rivers. The first and third communities presented similar values of richness (i.e. number of families) and diversity but the first was characterised by high abundance and was dominated by Baetidae (43.2%) while Hydropsychidae (36.3%) dominated the third community. The second community was dominated by Baetidae (33.4%), but it involved low abundance, richness and diversity samples and encompassed the microhabitats where no-macroinvertebrates were found. The performance of the MLP acknowledged the quality of the delineation and it indicated that the first community shows a clear affinity for microhabitats with aquatic vegetation and woody debris and the third for unshaded, fast flowing and shallow microhabitats on intermediate-sized substrate. Conversely, the second community occurred in deep and shaded microhabitats with low flow velocity and coarse substrate. We demonstrated that habitat structure and hydraulics are able to properly discriminate the macroinvertebrate communities, which, in turn, underlines their importance as drivers of community composition and abundance. Aquatic vegetation, woody debris, velocity and substrate index, followed by depth and shade, emerged as the most discriminant variables to understand macroinvertebrate communities in these tropical running waters. These results should enhance the implementation of ongoing and future EFA studies.

Text/Figure Separation in Document Images Using Docstrum Descriptor and Two-Level Clustering

We propose a novel algorithm for text/figure separation tailored for binary document images containing line drawings, block diagrams, charts, schemes and other kinds of business graphics. Most of the approaches for this task rely either on clever design of visual descriptor allowing to easily distinguish text and graphics regions or on the supervised learning using dataset of labeled text/figure regions. Such approaches often provide moderate separation accuracy when applied to document images which contain very diverse set of figure classes and lack sufficiently representative labeled training dataset. In contrast, our method is well-suited for vast variety of figure classes and capable of operating either in semi-supervised mode or unsupervised mode. We achieve this by leveraging unsupervised learning algorithms applied to Docstrum descriptors extracted from regions of interest and subsequent semi-supervised label propagation or unsupervised label inference. Another advantage of our method is its suitability for large scale data processing which is achieved through efficient kernel-approximating feature mapping applied to Docstrum descriptors and two-level clustering where fast mini-batch K-means algorithm is first applied to large scale data and only small number of resulting cluster centroids is subsequently processed by one of the more sophisticated clustering algorithms.

Machine Learning of Protein Interactions in Fungal Secretory Pathways.

In this paper we apply machine learning methods for predicting protein interactions in fungal secretion pathways. We assume an inter-species transfer setting, where training data is obtained from a single species and the objective is to predict protein interactions in other, related species. In our methodology, we combine several state of the art machine learning approaches, namely, multiple kernel learning (MKL), pairwise kernels and kernelized structured output prediction in the supervised graph inference framework. For MKL, we apply recently proposed centered kernel alignment and p-norm path following approaches to integrate several feature sets describing the proteins, demonstrating improved performance. For graph inference, we apply input-output kernel regression (IOKR) in supervised and semi-supervised modes as well as output kernel trees (OK3). In our experiments simulating increasing genetic distance, Input-Output Kernel Regression proved to be the most robust prediction approach. We also show that the MKL approaches improve the predictions compared to uniform combination of the kernels. We evaluate the methods on the task of predicting protein-protein-interactions in the secretion pathways in fungi, S.cerevisiae, baker’s yeast, being the source, T. reesei being the target of the inter-species transfer learning. We identify completely novel candidate secretion proteins conserved in filamentous fungi. These proteins could contribute to their unique secretion capabilities.

Dynamic evolving spiking neural networks for on-line spatio- and spectro-temporal pattern recognition

On-line learning and recognition of spatio- and spectro-temporal data (SSTD) is a very challenging task and an important one for the future development of autonomous machine learning systems with broad applications. Models based on spiking neural networks (SNN) have already proved their potential in capturing spatial and temporal data. One class of them, the evolving SNN (eSNN), uses a one-pass rank-order learning mechanism and a strategy to evolve a new spiking neuron and new connections to learn new patterns from incoming data. So far these networks have been mainly used for fast image and speech frame-based recognition. Alternative spike-time learning methods, such as Spike-Timing Dependent Plasticity (STDP) and its variant Spike Driven Synaptic Plasticity (SDSP), can also be used to learn spatio-temporal representations, but they usually require many iterations in an unsupervised or semi-supervised mode of learning. This paper introduces a new class of eSNN, dynamic eSNN, that utilise both rank-order learning and dynamic synapses to learn SSTD in a fast, on-line mode. The paper also introduces a new model called deSNN, that utilises rank-order learning and SDSP spike-time learning in unsupervised, supervised, or semi-supervised modes. The SDSP learning is used to evolve dynamically the network changing connection weights that capture spatio-temporal spike data clusters both during training and during recall. The new deSNN model is first illustrated on simple examples and then applied on two case study applications: (1) moving object recognition using address-event representation (AER) with data collected using a silicon retina device; (2) EEG SSTD recognition for brain–computer interfaces. The deSNN models resulted in a superior performance in terms of accuracy and speed when compared with other SNN models that use either rank-order or STDP learning. The reason is that the deSNN makes use of both the information contained in the order of the first input spikes (which information is explicitly present in input data streams and would be crucial to consider in some tasks) and of the information contained in the timing of the following spikes that is learned by the dynamic synapses as a whole spatio-temporal pattern.

Help-Training for semi-supervised support vector machines

In this paper, we propose to reinforce the Self-Training strategy in semi-supervised mode by using a generative classifier that may help to train the main discriminative classifier to label the unlabeled data. We call this semi-supervised strategy Help-Training and apply it to training kernel machine classifiers as support vector machines (SVMs) and as least squares support vector machines. In addition, we propose a model selection strategy for semi-supervised training. Experimental results on both artificial and real problems demonstrate that Help-Training outperforms significantly the standard Self-Training. Moreover, compared to other semi-supervised methods developed for SVMs, our Help-Training strategy often gives the lowest error rate.

Kerfdr: a semi-parametric kernel-based approach to local false discovery rate estimation

BackgroundThe use of current high-throughput genetic, genomic and post-genomic data leads to the simultaneous evaluation of a large number of statistical hypothesis and, at the same time, to the multiple-testing problem. As an alternative to the too conservative Family-Wise Error-Rate (FWER), the False Discovery Rate (FDR) has appeared for the last ten years as more appropriate to handle this problem. However one drawback of FDR is related to a given rejection region for the considered statistics, attributing the same value to those that are close to the boundary and those that are not. As a result, the local FDR has been recently proposed to quantify the specific probability for a given null hypothesis to be true.ResultsIn this context we present a semi-parametric approach based on kernel estimators which is applied to different high-throughput biological data such as patterns in DNA sequences, genes expression and genome-wide association studies.ConclusionThe proposed method has the practical advantages, over existing approaches, to consider complex heterogeneities in the alternative hypothesis, to take into account prior information (from an expert judgment or previous studies) by allowing a semi-supervised mode, and to deal with truncated distributions such as those obtained in Monte-Carlo simulations. This method has been implemented and is available through the R package kerfdr via the CRAN or at .

Semi-supervised discriminative classification with application to tumorous tissues segmentation of MR brain images

Due to the large data size of 3D MR brain images and the blurry boundary of the pathological tissues, tumor segmentation work is difficult. This paper introduces a discriminative classification algorithm for semi-automated segmentation of brain tumorous tissues. The classifier uses interactive hints to obtain models to classify normal and tumor tissues. A non-parametric Bayesian Gaussian random field in the semi-supervised mode is implemented. Our approach uses both labeled data and a subset of unlabeled data sampling from 2D/3D images for training the model. Fast algorithm is also developed. Experiments show that our approach produces satisfactory segmentation results comparing to the manually labeled results by experts.