Application of deep transfer learning for automated brain abnormality classification using MR images

Magnetic resonance imaging (MRI) is the most common imaging technique used to detect abnormal brain tumors. Traditionally, MRI images are analyzed manually by radiologists to detect the abnormal conditions in the brain. Manual interpretation of huge volume of images is time consuming and difficult. Hence, computer-based detection helps in accurate and fast diagnosis. In this study, we proposed an approach that uses deep transfer learning to automatically classify normal and abnormal brain MR images. Convolutional neural network (CNN) based ResNet34 model is used as a deep learning model. We have used current deep learning techniques such as data augmentation, optimal learning rate finder and fine-tuning to train the model. The proposed model achieved 5-fold classification accuracy of 100% on 613 MR images. Our developed system is ready to test on huge database and can assist the radiologists in their daily screening of MR images.     

Facial expression recognition in the wild,by fusion of deep learnt and hand-crafted features

In spite of the recent advancements in the field of deep learning based techniques for facial expression recognition, the efficiency of the state-of-the-art recognition methods in the wild scenarios, remains a challenge. The main reason behind the less efforts made for handling wild scenarios is two-folds: very less and varying levels of cues available to identify the distinguishable patterns of features (spatial and temporal) and non-availability of a big dataset to train a deep learning model. Recently, a huge dataset called AffectNet is introduced in the literature providing enough base to apply a deep learning model to train. This paper proposes an efficient combination of hand crafted and deep learning features for facial expression recognition in the wild. We use facial landmark points as hand-crafted features and XceptionNet for the deep learned features. We experiment with XceptionNet and Densenet propose the use of XceptionNet as it performs better compared to DenseNet, when applied on wild scenarios. The proposed fusion of the hand-crafted and XceptionNet features outperforms the state-of-the-art methods for facial expression recognition in the wild.     

Deep learning based liver cancer detection using watershed transform and Gaussian mixture model techniques

ObjectivesLiver cancer is one of the leading cause of death in all over the world. Detecting the cancer tissue manually is a difficult task and time consuming. Hence, a computer-aided diagnosis (CAD) is used in decision making process for accurate detection for appropriate therapy. Therefore the main objective of this work is to detect the liver cancer accurately using automated method.MethodsIn this work, we have proposed a new system called as watershed Gaussian based deep learning (WGDL) technique for effective delineate the cancer lesion in computed tomography (CT) images of the liver. A total of 225 images were used in this work to develop the proposed model. Initially, the liver was separated using marker controlled watershed segmentation process and finally the cancer affected lesion was segmented using the Gaussian mixture model (GMM) algorithm. After tumor segmentation, various texture features were extracted from the segmented region. These segmented features were fed to deep neural network (DNN) classifier for automated classification of three types of liver cancer i.e. hemangioma (HEM), hepatocellular carcinoma (HCC) and metastatic carcinoma (MET).ResultsWe have achieved a classification accuracy of 99.38%, Jaccard index of 98.18%, at 200 epochs using DNN classifier with a negligible validation loss of 0.062 during the classification process.ConclusionsOur developed system is ready to be tested with huge database and can aid the radiologist in detecting the liver cancer using CT images.     

Automatic skin lesions segmentation based on a new morphological approach via geodesic active contour

Automatic disease classification has been one of the most intensively searched in recent years due to the possibility of quickly providing a diagnosis to the patient. In this process, the segmentation of regions of interest of these diseases has a fundamental role in their subsequent classification. With skin lesions segmentation it is no different and in recent years many studies have achieved interesting results, becoming an important tool in aiding the medical diagnosis of skin diseases. In this work, a morphological geodesic active contour segmentation (MGAC) method is proposed with automatic initialization, using mathematical morphology which is a great partial differential equation approximation, with lower computational cost, no stability problems and fully automatic. The proposed method was tested in a stable and well-known dermoscopic images database provided by Pedro Hispano Hospital (PH2) and was compared with both methods that make use of machine learning or deep learning techniques such as fully convolutional networks (FCN), full resolution convolutional networks (FrCN), deep class-specific learning with probability based step-wise integration (DCL-PSI), and others, and also traditional methods like JSEG, statistical region merging (SRM), Level Set, ASLM and others. The MGAC showed good results in all similarity metrics compared in this work like Jaccard Index (86.16%), Dice coefficient (92.09%) and Matthew correlation coefficient (87.52%), and also achieves good results in sensitivity (91.72%), specificity (97.99%), accuracy (94.59%) and F-measure (93.82%). Thus, the proposed method presented better results in relation to all these metrics when compared to the traditional methods and still presented better results in relation to the methods that use machine learning or deep learning techniques in Jaccard Index, Dice coefficient and specificity. This confirm that the MGAC can efficiently segment skin lesions, presenting great potential to be applied in the aid of the medical diagnosis.     

Intelligent digital image firewall system for filtering privacy or sensitive images

As the commonest part of social networks, sharing images in social network not only provides more information, but also gives more intuitive view than text. However, images also can leak out information more easily than text, so the audit of image content is particularly essential. The disclosure of a tiny image, which involves sensitive information about individual, society even the state, may trigger a series of serious problems. In this paper, we design a kind of intelligent image firewall to detect and filter sensitive or privacy images. Two different approaches of the firewall are proposed. In the first approach, we propose an image firewall based on joint sparse representation, which can provide accurate and robust privacy prediction, and also can provide rich spatial relationship information. In the second approach, we propose a method based on the deep learning (Faster RCNN), which can predict the privacy relationships or actions (like kiss, hug and hand in hand) among the persons of an image. Experimental results show the effectiveness of the two kinds of approaches.     

Neurogenesis and the spacing effect: learning over time enhances memory and the survival of new neurons

Information that is spaced over time is better remembered than the same amount of information massed together. This phenomenon, known as the spacing effect, was explored with respect to its effect on learning and neurogenesis in the adult dentate gyrus of the hippocampal formation. Because the cells are generated over time and because learning enhances their survival, we hypothesized that training with spaced trials would rescue more new neurons from death than the same number of massed trials. In the first experiment, animals trained with spaced trials in the Morris water maze outperformed animals trained with massed trials, but there was not a direct effect of trial spacing on cell survival. Rather, animals that learned well retained more cells than animals that did not learn or learned poorly. Moreover, performance during acquisition correlated with the number of cells remaining in the dentate gyrus after training. In the second experiment, the time between blocks of trials was increased. Consequently, animals trained with spaced trials performed as well as those trained with massed, but remembered the location better two weeks later. The strength of that memory correlated with the number of new cells remaining in the hippocampus. Together, these data indicate that learning, and not mere exposure to training, enhances the survival of cells that are generated 1 wk before training. They also indicate that learning over an extended period of time induces a more persistent memory, which then relates to the number of cells that reside in the hippocampus.     

Neurogenesis and learning: acquisition and asymptotic performance predict how many new cells survive in the hippocampus

Previous research has shown that some associative learning tasks prevent the death of new neurons in the adult hippocampus. However, it is unclear whether it is mere exposure to the training stimuli that rescues neurons or whether successful learning of the task is required for enhanced neuronal survival. If learning is the important variable, then animals that learn better given the same amount of training should retain more of the new cells after learning than animals that do not learn as well. Here, we examined the effects of training versus learning on cell survival in the adult hippocampus. Animals were injected with BrdU to label a population of cells and trained one week later on one of two trace conditioning tasks, one of which depends on the hippocampus and one that does not. Increases in cell number occurred only in animals that acquired the learned response, irrespective of the task. There were significant correlations between acquisition and cell number, as well as between asymptotic performance and cell number. These data support the idea that learning and not simply training increases the survival of the new cells in the hippocampus.     

Looking in the wrong direction correlates with more accurate word learning

Previous research on lexical development has aimed to identify the factors that enable accurate initial word-referent mappings based on the assumption that the accuracy of initial word-referent associations is critical for word learning. The present study challenges this assumption. Adult English speakers learned an artificial language within a cross-situational learning paradigm. Visual fixation data were used to assess the direction of visual attention. Participants whose longest fixations in the initial trials fell more often on distracter images performed significantly better at test than participants whose longest fixations fell more often on referent images. Thus, inaccurate initial word-referent mappings may actually benefit learning.     

A deep model with combined losses for person re-identification

Person re-identification (PReID), which aims to re-identity a pedestrian from multiple non-overlapping cameras, has been significantly improved by deep learning system. There exist two popular deep frameworks used for PReID, i.e., identification and triplet models. Since these two frameworks have different loss functions, they have their own advantages and disadvantages. To combine the both advantages of two frameworks, in this paper, we propose using the triplet and Online Instance Matching (OIM) losses to train the carefully designed network. Given a triplet input images, the combined model can output the identities of the input images and learn a corresponding similarity measurement simultaneously. Experiments on CUHK01, CUHK03, Market-1501, and DukeMTMC-reID datasets demonstrate that the proposed model outperforms the compared state-of-the-art methods in most cases.     

Critical Appraisal of the Justification of Learning Outcome Design in the Subject Area of Religious Studies

Many studies have been carried out which suggest that students learn more effectively when introduced to teaching and learning objectives that promote deep learning over surface learning. Religious studies is a multi‐disciplinary subject concerned with promoting the study skills required for deep learning as these are innate to its approach to the wide variety of religious beliefs and practices found in the world. There are claims that learning outcomes curriculum design can promote the shift away from surface learning to a more deep approach. However, the promotion of learning outcomes can often originate from vocationalism where the specific requirements may not necessarily promote individual thought and independence. Religious studies is not perceived as a vocational subject area and needs therefore to examine very carefully the benefits of learning outcomes.     

In search of subtypes of Chinese developmental dyslexia

The dual-route model offers a popular way to classify developmental dyslexia into phonological and surface subtypes. The current study examined whether this dual-route model could provide a framework for understanding the varieties of Chinese developmental dyslexia. Three groups of Chinese children (dyslexics, chronological-age controls, and reading-level controls) were tested on Chinese exception character reading, pseudocharacter reading (analogous to English nonword reading), novel word learning, and some phonological and orthographic skills. It was found that Chinese exception character reading and pseudocharacter reading were highly correlated and that orthographic skills was a better predictor of both Chinese exception character and pseudocharacter reading than was phonological skills. More than half (62%) of the children in the dyslexia sample were classified as belonging to the surface subtype, but no children were classified as belonging to the phonological subtype. These results suggested that the lexical and sublexical routes in Chinese are highly interdependent or that there may be only one route from print to speech as suggested by the connectionist models. Chinese dyslexic children generally are characterized as having delays in various phonological and orthographic skills, but some, such as those identified as surface dyslexics in the current study, are more severely impaired.     

Self-regulatory Behaviors and Approaches to Learning of Arts Students: A Comparison Between Professional Training and English Learning

This study investigated the self-regulatory behaviors of arts students, namely memory strategy, goal-setting, self-evaluation, seeking assistance, environmental structuring, learning responsibility, and planning and organizing. We also explored approaches to learning, including deep approach (DA) and surface approach (SA), in a comparison between students’ professional training and English learning. The participants consisted of 344 arts majors. The Academic Self-Regulation Questionnaire and the Revised Learning Process Questionnaire were adopted to examine students’ self-regulatory behaviors and their approaches to learning. The results show that a positive and significant correlation was found in students’ self-regulatory behaviors between professional training and English learning. The results indicated that increases in using self-regulatory behaviors in professional training were associated with increases in applying self-regulatory behaviors in learning English. Seeking assistance, self-evaluation, and planning and organizing were significant predictors for learning English. In addition, arts students used the deep approach more often than the surface approach in both their professional training and English learning. A positive correlation was found in DA, whereas a negative correlation was shown in SA between students’ self-regulatory behaviors and their approaches to learning. Students with high self-regulation adopted a deep approach, and they applied the surface approach less in professional training and English learning. In addition, a SEM model confirmed that DA had a positive influence; however, SA had a negative influence on self-regulatory behaviors.     

An effective daily box office prediction model based on deep neural networks

The task of the daily box office prediction model is to build a dynamic prediction model to rolling forecast daily box office. It is a complex task as the movie box office has a short life cycle, and the static data and dynamic data that affect the trend of box office are heterogeneous. This paper proposes an end-to-end deep learning model for daily box office prediction, called Deep-DBP which consists of temporal component and static characteristics component. The temporal component is the main component which uses LSTM to learn the temporal dependencies between data points. The static characteristics component is an auxiliary component and it integrates static characteristics to improve prediction effect. The Deep-DBP can overcome the problems that the ARIMA and traditional ANN model cannot solve. The structure of input and output proposed in the model can well handle short time series prediction problem. It is a successful case in dealing with multi-source and multi-view data, addition of static characteristics component reduces the prediction error by 7%. The prediction error of Deep-DBP is 30.1%, which is better than that of the previous model. The experiment proved that the more training data collected, the better the prediction effect.     

Learning representation hierarchies by sharing visual features: a computational investigation of Persian character recognition with unsupervised deep learning

In humans, efficient recognition of written symbols is thought to rely on a hierarchical processing system, where simple features are progressively combined into more abstract, high-level representations. Here, we present a computational model of Persian character recognition based on deep belief networks, where increasingly more complex visual features emerge in a completely unsupervised manner by fitting a hierarchical generative model to the sensory data. Crucially, high-level internal representations emerging from unsupervised deep learning can be easily read out by a linear classifier, achieving state-of-the-art recognition accuracy. Furthermore, we tested the hypothesis that handwritten digits and letters share many common visual features: A generative model that captures the statistical structure of the letters distribution should therefore also support the recognition of written digits. To this aim, deep networks trained on Persian letters were used to build high-level representations of Persian digits, which were indeed read out with high accuracy. Our simulations show that complex visual features, such as those mediating the identification of Persian symbols, can emerge from unsupervised learning in multilayered neural networks and can support knowledge transfer across related domains.     

Modeling the development of vehicle lateral control skills in a cognitive architecture

The development of lateral control skills is crucial to driving safety. The current study examined a computational method using a cognitive architecture to model the learning process of vehicle lateral control. In a fixed-base driving simulator, an experiment compared the lateral control performance of non-drivers, novices, and experienced drivers. A cognitive model using Adaptive Control of Thought-Rational (ACT-R) was built to model the learning process of lateral control skills. The modeling results were compared with the human results. The drivers with more experience had better lateral control performance. The model produced similar results as the human results and modeled the progress of learning. The model provided a computational explanation for the mechanisms of lateral control skill learning. Implication and future studies were discussed.     

An intelligent facial expression recognition system with emotion intensity classification

Facial expressions play a crucial role in emotion recognition as compared to other modalities. In this work, an integrated network, which is capable of recognizing emotion intensity levels from facial images in real time using deep learning technique is proposed. The cognitive study of facial expressions based on expression intensity levels are useful in applications such as healthcare, coboting, Industry 4.0 etc. This work proposes to augment emotion recognition with 2 other important parameters, valence and emotion intensity. This helps in better automated responses by a machine to an emotion. The valence model helps in classifying emotion as positive and negative emotions and discrete model classifies emotions as happy, anger, disgust, surprise and neutral state using Convolution Neural Network (CNN). Feature extraction and classification are carried out using CMU Multi-PIE database. The proposed architecture achieves 99.1% and 99.11% accuracy for valence model and discrete model respectively for offline image data with 5-fold cross validation. The average accuracy achieved in real time for valance model and discrete model is 95% & 95.6% respectively. Also, this work contributes to build a new database using facial landmarks, with three intensity levels of facial expressions which helps to classify expressions into low, mild and high intensities. The performance is also tested for different classifiers. The proposed integrated system is configured for real time Human Robot Interaction (HRI) applications on a test bed consisting of Raspberry Pi and RPA platform to assess its performance.     

Earthquake disaster avoidance learning system using deep learning

The popularity of deep learning has influenced the field of surveillance and human safety. We adopt the advantages of deep learning techniques to recognize potentially harmful objects inside living rooms, offices, and dining rooms during earthquakes. In this study, we propose an educational system to teach earthquake risks using indoor object recognition based on deep learning algorithms. The system is based on the You Look Only Once (YOLO) deployed on our cloud-based server named Earthquake Situation Learning System (ESLS) for the detection of harmful objects associated with risk tags. ESLS is trained on our own indoor images dataset. The user interacts with the ESLS server through video or image files, and the object detection algorithm using YOLO recognizes the indoor objects with associated risk tags. Results show that the service time of ESLS is low enough to serve it to users in 0.8 s on average, including processing and communication times. Furthermore, the accuracy of the harmful object detection is 96% in the general indoor lighting situation. The results show that the proposed ESLS is applicable to real service for teaching the earthquake disaster avoidance.     

Convolutional neural network based Alzheimer’s disease classification from magnetic resonance brain images

Alzheimer’s disease, the most common form of dementia is a neurodegenerative brain order that has currently no cure for it. Hence, early diagnosis of such disease using computer-aided systems is a subject of great importance and extensive research amongst researchers. Nowadays, deep learning or particularly convolutional neural network (CNN) is getting more attention due to its state-of-the-art performances in variety of computer vision tasks such as visual object classification, detection and segmentation. Several recent studies, that have used brain MRI scans and deep learning have shown promising results for diagnosis of Alzheimer’s disease. However, most common issue with deep learning architectures such as CNN is that they require large amount of data for training. In this paper, a mathematical model P_FS_EC_TL based on transfer learning is used in which a CNN architecture, VGG-16 trained on ImageNet dataset is used as a feature extractor for the classification task. Experimentation is performed on data collected from Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The accuracy of the 3-way classification using the described method is 95.73% for the validation set.     

发展性阅读障碍亚类型研究进展

发展性阅读障碍是一种特殊的学习障碍,伴有多种认知缺陷并且存在不同的亚类型。依据相关的阅读模型理论,阅读障碍可划分为语音型和表层型。从认知缺陷角度出发,语音加工缺陷是主要的缺陷表现,以此为特征形成一种主要的阅读障碍的亚类型,同时还有以正字法加工缺陷和快速命名缺陷为主的其他亚类型。而以基本感知觉缺陷为标准,主要有以视觉加工缺陷和以听觉加工缺陷为主的两种亚类型。在汉语条件下,依据同样的阅读模型理论,语音型阅读障碍亚类型比例明显低于拼音文字条件下的。汉语阅读障碍也具有分别以语音加工缺陷、快速命名缺陷和正字法加工缺陷为主要认知缺陷的亚类型。未来有必要从神经机制角度进一步明确不同亚类型的神经基础。