Search

5 Tables and 3 Figures from Integrated Array-Comparative Genomic Hybridization and Expression Array Profiles Identify Clinically Relevant Molecular Subtypes of Glioblastoma

Glioblastoma, the most aggressive primary brain tumor in humans, exhibits a large degree of molecular heterogeneity. Understanding the molecular pathology of a tumor and its linkage to behavior is an important foundation for developing and evaluating approaches to clinical management. Here we integrate array-comparative genomic hybridization and array-based gene expression profiles to identify relationships between DNA copy number aberrations, gene expression alterations, and survival in 34 patients with glioblastoma. Unsupervised clustering on either profile resulted in similar groups of patients, and groups defined by either method were associated with survival. The high concordance between these separate molecular classifications suggested a strong association between alterations on the DNA and RNA levels. We therefore investigated relationships between DNA copy number and gene expression changes. Loss of chromosome 10, a predominant genetic change, was associated not only with changes in the expression of genes located on chromosome 10 but also with genome-wide differences in gene expression. We found that CHI3L1/YKL-40 was significantly associated with both chromosome 10 copy number loss and poorer survival. Immortalized human astrocytes stably transfected with CHI3L1/YKL-40 exhibited changes in gene expression similar to patterns observed in human tumors and conferred radioresistance and increased invasion in vitro. Taken together, the results indicate that integrating DNA and mRNA-based tumor profiles offers the potential for a clinically relevant classification more robust than either method alone and provides a basis for identifying genes important in glioma pathogenesis.

Skin cancers are one of the most common cancers in the world. Early detections and treatments of skin cancers can greatly improve the survival rates of patients. In this paper, a skin lesions classification system is developed with deep convolutional neural networks of ResNet50, which may help dermatologists to recognize skin cancers earlier. We utilize the ResNet50 as a pre-trained model. Then, by transfer learning, it is trained on our skin lesions dataset. Image preprocessing and dataset balancing methods are used to increase the accuracy of the classification model. In classification of skin diseases, our model achieves an overall accuracy of 83.74% on nine-class skin lesions. The experimental results show an impressive effect of the ResNet50 model in finegrained skin lesions classification and skin cancers recognition.

Background Nevus is very common; however, melanoma is slightly related to the deterioration of nevus because of its vulnerability to solarization, friction, aging, heredity, and other factors. Early diagnosis is essential for melanoma treatment, since patients have a high survival rate with early detection and treatment. Computer-aided diagnosis has been applied in the differential diagnosis of melanoma and benign nevi and achieved high accuracy, but it does not suit the screening of nevi because most studies are based on dermoscopy with a narrow field of vision and performed by professional doctors. Therefore, this study aimed to present the accuracy and effectiveness of our algorithm. Methods Based on whole-face images of patients, the authors used logistic regression and the Newton method to detect the nevus region. Then, Python and OpenCV were employed to detect the lesion edge and compute the area of the regions. A multicenter clinical trial with a sample size of 600 was then conducted to evaluate the effectiveness of the algorithm. Results The algorithm detected 2672 nevi from 600 patients, in which there were 195 patients of missed diagnosis and 310 patients of misdiagnosis. The Kappa value between 2 groups was 0.860 (>0.8). Paired t-test showed no significant difference between 2 groups' area results (P = 0.265, P > 0.05). Conclusion Within the limitations of this study, the authors demonstrated a high agreement between algorithm's detection and doctor's diagnosis. Our new algorithm has great effectiveness in nevus detection, edge segmentation, and area measurement.

Deep Neural Networks (DNNs) have achieved high accuracy on image classification. However, a small disturbance to an input may fool the networks to misclassify the label, which can cause a series of security and social problems. Thus, the robustness of DNNs must be ensured, particularly to those safety-critical systems. In this paper, we focus on the problem of measuring the robustness of ReLU-based DNNs, which can be equivalently formulated to solve a Mixed Integer Linear Programming problem (MILP). The complexity of solving MILP is directly related to the number of integer variables. We propose an efficient method for robustness measurement and verification by pruning the search space of MILP problems. Particularly, we design a greedy algorithm based on linear programming (LP) to determine the reasonable boundary. Then the search space is pruned by setting the boundary to integer variables in MILP. The comparison experiments on five classifiers trained on MNIST and CIFAR-10 datasets show our method outperforms other related tools in terms of efficiency and accuracy.

In this study, we propose a novel Joint-HSMM method that combines the CNN and probabilistic models (HSMM) for heart sound segmentation. First, we build a semantic segmentation model and a time series probability model based on U-Net and HSMM respectively, and fusion model results by average consistency. The proposed method are conducted on various datasets, including the publicly available PhysioNet and a children heart sound dataset collected by Shanghai Xinhua Hospital. For the second dataset, we make use of Transfer Learning techniques for early convergence and robustness. It is shown that our method has outperformed the current state-of-the-art method by achieving the average recall, F1-score of 96.29%, 96.14% on PhysioNet dataset, and 86.95% and 87.24% on children heart sound dataset respectively.

Cloze test is a common test in language examinations. It is also a research direction of natural language processing, which is an important field of artificial intelligence. In general, some words in a complete article are hidden, and several candidates are given to let the student choose the correct hidden word. To explore whether machine can do cloze test, we have done some research to build down-stream tasks of BERT for cloze test. In this paper, we consider the compound words in articles and make an improvement to help the model handling these kind of words. The experimental results show that our model performs well on questions of compound words and has better accuracy on CLOTH dataset.

In recent years, crowdsourced testing, which uses collective intelligence to solve complex software testing tasks has gained widespread attention in academia and industry. However, due to a large number of workers participating in crowdsourced testing tasks, the submitted test reports set is too large, making it difficult for developers to review test reports. Therefore, how to effectively process and integrate crowdsourced test reports is always a significant challenge in the crowdsourced testing process. This paper deals with the crowdsourced test reports processing, sorts out some achievements in this field in recent years, and classifies, summarizes, and compares existing research results from four directions: duplicated reports detection, test reports aggregation and classification, priority ranking, and reports summarization. Finally explored the possible research directions, opportunities and challenges of the crowdsourced test reports.

Recently, mobile devices are more popular than computers. However, mobile apps are not as thoroughly tested as desktop ones, especially for graphical user interface (GUI). In this paper, we study the detection and segmentation of graphical elements on GUIs for mobile apps based on deep learning. It is the preliminary work of GUI testing for mobile apps based on artificial intelligence. We create a dataset, which consists of 2,100 GUI screenshots (or pages) labeled with 42,156 graphic elements in 8 classes. Based on our dataset, we adopt Mask R-CNN to train the detection and segmentation of graphic elements on GUI screenshots. The experimental results show that the mAP value achieves 98%.

Imbalance data is a common problem in machine learning task, which often impacts the accuracy of models. An effective way to solve it is to increase the number of minority class samples in the dataset. Many methods are put forward to solve the problem of imbalance data in machine learning. But these are all for low-dimensional data. For high-dimensional data, such as images, these methods are not well applicable. In this paper, an image generation method based on generative adversarial network is introduced to do pattern learning for samples of minority class in the dataset, so as to realize the expansion of data for minority class. And finally the classification networks for skin lesions are trained by data collaboration which consist of real images and generated images. The experimental results indicate that the accuracy of networks are further improved by the addition of generated images while alleviating the imbalance problem to some extent.

Convolutional neural networks have made unprecedented breakthroughs in various tasks of computer vision. Due to its complex nonlinear model structure and the high latitude and complexity of data distribution, it has been criticized as an unexplained “black box”. Therefore, explaining the neural network model and uncovering the veil of the neural network have become the focus of attention. This paper starts with the term “interpretability”, summarizes the results of the interpretability of convolutional neural networks in the past three years (2016–2018), and analyses them with interpretable methods. Firstly, the concept of “interpretability” is introduced. Then the existing research achievements are classified and compared from four aspects, data characteristics and rule processing, model internal spatial analysis, interpretation and prediction, and model interpretation. Finally pointed out the possible research directions.

In this paper, a binary segmentation method HCDS based on salient object detection is proposed to solve the segmentation of lesions in skin images. The original U-Net model encounters some problems in its application to skin image segmentation, one of which is the non-pathological region information remained in the segmentation results. Our HCDS method improves its basic structure by adding a hybrid convolution module to the direct connection between the down-sampling and up-sampling. It deepens the abstract understanding of the input features in shallow layers and further eliminate the information interference of the non-pathological regions. Besides, the method employs a deeply supervised structure in each stage of up-sampling to learn from the output feature and ground truth. Finally, the HCDS integrates the multi-path outputs to obtain the best result. The experimental results show that the proposed method HCDS can effectively segment the lesion regions from the skin images. Our evaluation metrics about the HCDS are obviously superior to the original U-Net's.