Author : Aakriti Adhikari
Publisher :
ISBN 13 :
Total Pages : 116 pages
Book Rating : 4.:/5 (126 download)
Book Synopsis Skin Cancer Detection Using Generative Adversarial Network and an Ensemble of Deep Convolutional Neural Networks by : Aakriti Adhikari
Download or read book Skin Cancer Detection Using Generative Adversarial Network and an Ensemble of Deep Convolutional Neural Networks written by Aakriti Adhikari and published by . This book was released on 2019 with total page 116 pages. Available in PDF, EPUB and Kindle. Book excerpt: Over the past few years, Deep learning has been widely used in medical imaging for classification and segmentation and has been successful in providing better diagnostic accuracy. The state-of- the-art deep learning algorithms are built using neural networks arranged in layers where the first layer extracts basic information of images like edges, colors etc so that the output of one layer is fed as input to the next consecutive layers. Thus, increasing the complexity of learning with increase of layers. In comparison with the traditional machine learning algorithms, deep learning has many advantages and is an automatic process. However, it requires large scale annotated data for better performance and is thus constrained by limited size of available public datasets. To overcome data constraints, this thesis proposes a more efficient and novel scheme comprising techniques that use an ensemble of convolution neural networks (CNN) with generative adversarial network (GAN) based augmentation to improve the diagnostic accuracy. Also, the implementation of the proposed technique on skin lesion dataset of limited size to classify melanoma is presented. In this work, images are augmented to increase the dataset size using two different augmentation techniques: traditional augmentation and GAN based augmentation. GAN based augmentation is based on neural networks, where, two neural networks compete against each other to produce visually realistic synthetic images. For the classification, CNN and an ensemble of CNNs are trained on the final enlarged training dataset comprising original images and synthetic images. Here, the ensemble of CNN techniques combines five trained CNNs into a single meta-classifier. One hundred and ninety-three skin lesion test images were used to validate our proposed methods. The first method proposed in this research incorporated traditional to enlarge the training dataset and trained a CNN classifier to classify skin cancer into malignant or benign. The performance of the model on the test dataset was observed with 79.29% of accuracy, 77% sensitivity and 81.38% specificity. The second proposed method incorporated both traditional augmentation and GAN based augmentation to enlarge the training dataset and a CNN classifier was trained in a similar way as compared to the first method. This method was validated with 81.32% accuracy, 80.26% sensitivity and 82.78% specificity. The final proposed method utilized both of these augmentation techniques to enlarge the training dataset and an ensemble of five CNN classifiers was trained for the classification. It performed better with the performance accuracy of 84.30% , sensitivity of 84.32% and specificity of 84.21% . Further, it prevented trapping into local maxima, causing performance to increase. The scheme with the proposed methods can be a better choice while dealing with the limited training dataset.
