摘要 :
Neural Architecture Search (NAS) is one of the most recent developments in automating the design process for deep convolutional neural network (DCNN) architectures. NAS and later Efficient NAS (ENAS) based models have been adopted...
展开
Neural Architecture Search (NAS) is one of the most recent developments in automating the design process for deep convolutional neural network (DCNN) architectures. NAS and later Efficient NAS (ENAS) based models have been adopted successfully for various applications including ultrasound image classification for breast lesions. Such a data driven approach leads to creation of DCNN models that are more applicable to the data set at hand but with a risk for model overfitting. In this paper, we first investigate the extent of the ENAS model generalization error problem by using different test data sets of ultrasound images of breast lesions. We have observed a significant reduction of overall average accuracy by nearly 10% and even more severe reduction of specificity rate by more than 20%, indicating that model generalization error is a serious issue with ENAS models for breast lesion classification in ultrasound images. To overcome the generalization error, we examined the effectiveness of a range of techniques including reducing model complexity, use of data augmentation, and use of unbalanced training sets. Experimental results show that different methods for the tuned ENAS models achieved different levels of accuracy when they are tested on internal and two external test data sets. The paper demonstrates that ENAS models trained on an unbalanced dataset with more benign cases tend to generalize well on unseen images achieving average accuracies of 85.8%, 82.7%, and 88.1 % respectively for the internal and the two external test data sets not only on specificity alone, but also sensitivity. In particular, the generalization of the refined models across internal and external test data is maintained.
收起
摘要 :
Neural Architecture Search (NAS) is one of the most recent developments in automating the design process for deep convolutional neural network (DCNN) architectures. NAS and later Efficient NAS (ENAS) based models have been adopted...
展开
Neural Architecture Search (NAS) is one of the most recent developments in automating the design process for deep convolutional neural network (DCNN) architectures. NAS and later Efficient NAS (ENAS) based models have been adopted successfully for various applications including ultrasound image classification for breast lesions. Such a data driven approach leads to creation of DCNN models that are more applicable to the data set at hand but with a risk for model overfitting. In this paper, we first investigate the extent of the ENAS model generalization error problem by using different test data sets of ultrasound images of breast lesions. We have observed a significant reduction of overall average accuracy by nearly 10% and even more severe reduction of specificity rate by more than 20%, indicating that model generalization error is a serious issue with ENAS models for breast lesion classification in ultrasound images. To overcome the generalization error, we examined the effectiveness of a range of techniques including reducing model complexity, use of data augmentation, and use of unbalanced training sets. Experimental results show that different methods for the tuned ENAS models achieved different levels of accuracy when they are tested on internal and two external test data sets. The paper demonstrates that ENAS models trained on an unbalanced dataset with more benign cases tend to generalize well on unseen images achieving average accuracies of 85.8%, 82.7%, and 88.1% respectively for the internal and the two external test data sets not only on specificity alone, but also sensitivity. In particular, the generalization of the refined models across internal and external test data is maintained.
收起
摘要 :
Manual inspections of glass facade of high rising buildings are expensive, time-consuming and potentially life-threatening for both inspectors and pedestrians on the street. Advances in machine learning for image/video analysis an...
展开
Manual inspections of glass facade of high rising buildings are expensive, time-consuming and potentially life-threatening for both inspectors and pedestrians on the street. Advances in machine learning for image/video analysis and availability of affordable unmanned aerial vehicles (UAVs) with onboard video recording and processing sensors provide opportunities for smart, safe and automatic glass facade inspections. This paper is concerned with developing an effective solution for recognizing cracked glass panels, which can be installed on board a UAV. From static 2D photographic images, the proposed solution analyzes textural patterns of smooth glass surface and crack segments, linearity of detected crack segments, geometrical characteristics of crack curvatures and the crack pixel patterns, captures these discriminative features for glass cracks using Uniform Local Binary Pattern (ULBP), histograms of linearity, geometrical curvature descriptors with fixed length connected pixel configurations, and accordingly classifies images of cracked and non-cracked glass panels using a kNN classifier. Experimental results with images of different resolutions acquired by a UAV drone in a real office building setting and images collected through Google search demonstrate that the proposed solution achieves promising results with accuracy rates in excess of 80% and even as high as 91% despite the presence of reflections.
收起
摘要 :
Manual inspections of glass facade of high rising buildings are expensive, time-consuming and potentially life-threatening for both inspectors and pedestrians on the street. Advances in machine learning for image/video analysis an...
展开
Manual inspections of glass facade of high rising buildings are expensive, time-consuming and potentially life-threatening for both inspectors and pedestrians on the street. Advances in machine learning for image/video analysis and availability of affordable unmanned aerial vehicles (UAVs) with onboard video recording and processing sensors provide opportunities for smart, safe and automatic glass facade inspections. This paper is concerned with developing an effective solution for recognizing cracked glass panels, which can be installed on board a UAV. From static 2D photographic images, the proposed solution analyzes textural patterns of smooth glass surface and crack segments, linearity of detected crack segments, geometrical characteristics of crack curvatures and the crack pixel patterns, captures these discriminative features for glass cracks using Uniform Local Binary Pattern (ULBP), histograms of linearity, geometrical curvature descriptors with fixed length connected pixel configurations, and accordingly classifies images of cracked and non-cracked glass panels using a kNN classifier. Experimental results with images of different resolutions acquired by a UAV drone in a real office building setting and images collected through Google search demonstrate that the proposed solution achieves promising results with accuracy rates in excess of 80% and even as high as 91% despite the presence of reflections.
收起
摘要 :
Ultrasound scan (US) imagery is an important tool for radiologists to make a fast and reliable diagnosis decision about breast lesion status (benign or malignant). Accurate and automatic segmentation of breast lesion is critical f...
展开
Ultrasound scan (US) imagery is an important tool for radiologists to make a fast and reliable diagnosis decision about breast lesion status (benign or malignant). Accurate and automatic segmentation of breast lesion is critical for annotating the lesion characteristics such as margin smoothness and regularity in support of the diagnosis decision. Fully convolutional network (FCN) is one of the commonly used deep learning neural network methods for semantic segmentation. This paper is concerned with effective adaptation of the FCN solutions for segmenting breast lesions from 2D ultrasound images. The paper aims to first evaluate the existing FCN solution for solving the problem at hand and compare its performance with another popular method using U-Net. The paper then highlights one key issue with the FCN, i.e. false positive pixels near the boundary of a lesion and false positive pixels forming false lesions. The paper then investigates several methods in reducing such false positive pixels, including the use of data augmentation in training the classification model and use of loss functions in training the models. Experimental results using several data sets collected from various sources show that our adapted FCN method outperforms U-Net-based solutions in general and the false positive reduction methods we attempted have reduced the false positive pixels in both regions close to lesion boundary and separate from true lesion regions.
收起
摘要 :
The border irregularity of lesions or tumors is an important sign (or feature) contributing to the prediction of the tumor malignancy. This paper is concerned with developing automatic computer vision methods for assessing and rec...
展开
The border irregularity of lesions or tumors is an important sign (or feature) contributing to the prediction of the tumor malignancy. This paper is concerned with developing automatic computer vision methods for assessing and recognizing thyroid nodule border irregularity from ultrasound images. Unlike many existing schemes, our methods rely on a small set of points on the nodule border marked manually by clinicians. To mitigate the absence of a fully segmented lesion boundary, we first apply the cubic-spline interpolation of the region of interest (ROI) points to approximate the lesion border and then select equal numbers of points from the approximated border using equal angular distances. We developed two complementary approaches to investigate the global (big indentations and protrusions) and local (small zigzag) irregularity features of the nodule. The first approach includes two Euclidian distances-based methods and a method inspired by Fractal Dimensions (FD). The distances-based methods facilitate the use of the interpolated border and their radial distance functions measured from ROI points to a reference point (centroid) or reference shape (Convex hull), while the FD inspired method uses interpolated border and a fitted ellipse perimeter ratio to calculate an irregularity index. The second approach facilitates the texture analysis within the constructed ribbons around the border line of different widths using feature vector of uniform local binary pattern (ULBP). We evaluate and compare the performance of our methods from the two approaches by using two datasets consisting of 395 and 100 ultrasound images of thyroid nodules collected from two hospitals and labelled by experienced radiologists respectively. The first is used as training and internal testing set, while the second is used for external testing. We shall show the viability of our methods attaining accuracy rates between 70% and 90%.
收起
摘要 :
Convolutional neural networks have shown outstanding object recognition performance, especially for visual recognition tasks such as tumor classification in 2D ultrasound (US) images. In Computer-Aided Diagnosis (CAD) systems, int...
展开
Convolutional neural networks have shown outstanding object recognition performance, especially for visual recognition tasks such as tumor classification in 2D ultrasound (US) images. In Computer-Aided Diagnosis (CAD) systems, interpreting CNN's decision is crucial for accepting the system in the clinical use. This paper is concerned with 'visual explanations' for decisions from CNN models trained on ultrasound images. In particular, we investigate the link between the CNN decision and the calcification cancer sign in breast lesion classification task. To this end, we study the output visualization of two different breast lesion recognition CNN models in two folds: Firstly, we explore two existing visualization approaches, Grad-CAM and CRM, to gain insight into the function of feature layers. Secondly, we introduce an adaptive Grad-CAM, called EGrad-CAM, which uses information entropy to freeze feature maps with no or minimal information. Extensive analysis and experiments using 1624 US images and two breast classification models show that calcification feature contributes to the CNN classification decision for both malignant and benign lesions. Furthermore, we show many feature maps in the final convolution layer are not contributing to the CNN decision, and our EGrad-CAM produces similar visualization output to Grad-CAM using 24%-87% of the feature maps. Our study demonstrates that the CNN decision visualization is a promising direction for bridging the gap between CNN classification decision of US images of breast lesions and cancer signs.
收起
摘要 :
Convolutional neural networks have shown outstanding object recognition performance, especially for visual recognition tasks such as tumor classification in 2D ultrasound (US) images. In Computer-Aided Diagnosis (CAD) systems, int...
展开
Convolutional neural networks have shown outstanding object recognition performance, especially for visual recognition tasks such as tumor classification in 2D ultrasound (US) images. In Computer-Aided Diagnosis (CAD) systems, interpreting CNN's decision is crucial for accepting the system in the clinical use. This paper is concerned with 'visual explanations' for decisions from CNN models trained on ultrasound images. In particular, we investigate the link between the CNN decision and the calcification cancer sign in breast lesion classification task. To this end, we study the output visualization of two different breast lesion recognition CNN models in two folds: Firstly, we explore two existing visualization approaches, Grad-CAM and CRM, to gain insight into the function of feature layers. Secondly, we introduce an adaptive Grad-CAM, called EGrad-CAM, which uses information entropy to freeze feature maps with no or minimal information. Extensive analysis and experiments using 1624 US images and two breast classification models show that calcification feature contributes to the CNN classification decision for both malignant and benign lesions. Furthermore, we show many feature maps in the final convolution layer are not contributing to the CNN decision, and our EGrad-CAM produces similar visualization output to Grad-CAM using 24%-87% of the feature maps. Our study demonstrates that the CNN decision visualization is a promising direction for bridging the gap between CNN classification decision of US images of breast lesions and cancer signs.
收起
摘要 :
Data mining has been introduced into computing curricula. A data mining module should emphasise not only the technical but also the practical sides of the subject. This paper stresses the importance of using a data mining project ...
展开
Data mining has been introduced into computing curricula. A data mining module should emphasise not only the technical but also the practical sides of the subject. This paper stresses the importance of using a data mining project as a critical element of the course work. The paper outlines the intended learning outcomes and the expectations from students. The paper proposes a framework for project administration and assessment. By using a number of past projects as case studies, the paper demonstrates the project work involved and summarises good and bad experiences in running the project. The paper highlights the uncertain nature of data mining and consequent challenges and difficulties. The paper is intended to contribute towards a wider debate over the best practices in teaching, learning and assessment of data mining.
收起
摘要 :
Data mining has been introduced into computing curricula. A data mining module should emphasise not only the technical but also the practical sides of the subject. This paper stresses the importance of using a data mining project ...
展开
Data mining has been introduced into computing curricula. A data mining module should emphasise not only the technical but also the practical sides of the subject. This paper stresses the importance of using a data mining project as a critical element of the coursework. The paper outlines the intended learning outcomes and the expectations from students. The paper proposes a framework for project administration and assessment. By using a number of past projects as case studies, the paper demonstrates the project work involved and summarises good and bad experiences in running the project. The paper highlights the uncertain nature of data mining and consequent challenges and difficulties. The paper is intended to contribute towards a wider debate over the best practices in teaching, learning and assessment of data mining.
收起