Skin lesion classification using deep convolutional neural network and HSV color space

In this work, deep convolutional neural network and HSV color space based skin lesion classification using ISIC 2017 challenge dataset was analyzed. The main aim of the challenge is to assist participants to develop analytical image tools in order to automatically make a diagnosis of melanoma from dermoscopic images. Skin lesions image analysis consists of three parts: lesion segmentation, detection and localization and disease classification. In this study, it was concentrated on skin lesion classification part of the challenge. This dataset consists of training data (2000 images) to participate in all image analysis. A separate public verification data set (150 images) and a blind held out a test dataset (600 images) is provided to generate verification and testing results. Two types of classification tests are performed. First one is the classification of benign or malignant images. The second one is two independent binary image classification tasks that involve three unique diagnoses of skin lesions (melanoma, nevus, and seborrheic keratosis). In the first binary classification task, it is asked to distinguish between melanoma and nevus + seborrheic keratosis images. In the second binary classification task, it is asked to distinguish between seborrheic keratosis and nevus + melanoma images. All initial images in data sets include images of segmentation of lesion which are accompanied by an expert manual trace of borders of defeat in a type of a binary mask. Masks are created by the specialist clinical physician with use of semi-automatic process (with use of the initial point provided by the user, an algorithm of filling of floods which is adjusted by the user and morphological filtration) or manual process (from a series of the polyline of the point provided by the user). Binary masks help us to find the bounding rectangles of the skin lesion area. An easy way to detect the ROI is to use of OpenCV library that has contour detection function. It was applied to images with a binary mask resulting in coordinates of the bounding boxes. Using these rectangles and keeping the aspect ratio the same, lesion areas have been segmented and resized to 500x500 pixel in size. Before submission of images in a deep convolutional neural network, the images of RGB are transformed to the HSV color space. In the RGB color space when the value of pixel changes, changes the intensity of red, green and blue channels of this color pixel. It means that color, the intensity, and saturation of pixel are connected with variations of color. Observation of change of color in the difficult color environment becomes difficult; however the HSV color space divides color into a hue, saturation, and value. In other words it means that observation of color variations can be distinguished individually. Thus, the HSV color space describes color better, than the RGB color space. After preprocessing data augmentation was performed, such as rotation, shifting and flipping of images. This technique was utilized since our data size is relatively small to train the neural network. Using data augmentation images with the same context but shifted, rotated and flipped were produced. It provides different points of view on images. The first test that includes classification of benign and malignant images, transfer learning was used. As a model InceptionV3 was chosen. This network model is produced by Google. The second test consists of two subtasks: classification of (A) melanoma and nevus + seborrheic keratosis; and (B) seborrheic keratosis and nevus + melanoma images. For these tasks, a deep network with batch normalization between each convolutional layers was built. Batch normalization helps to increase the learning speed. As performance metrics for evaluation of all trained models on a test set, the area under curve and accuracy metrics were used. For the first (benign/malignant) classification, 0.55 area under the ROC curve (AUC) and 70% accuracy on the test set were achieved. The second one contains two tasks: for (melanoma/nevus+seborrheic keratosis) task, 0.61 area under the ROC curve and 80% accuracy and for (seborrheic keratosis/nevus+melanoma) task 0.65 area under the ROC curve and 87% accuracy were reached. All works were performed on the notebook, which includes a CPU of the Intel i5 6th generation and a 2 GB Nvidia GeForce 940M GPU. Since a deep convolutional neural network requires computing power, NVIDIA, the CUDA toolkit and cuDNN have been used. They make the design environment accessible in order to create highly efficient GPU-accelerated applications. In addition, TensorFlow and Keras were preferred as high-level frameworks for rapid prototype development of the study. TensorFlow is a public library of software for machine learning developed by Google to solve the problems of building and learning a neural network, to automatically find and classify images, having reached the quality of human perception. Keras is an open neural network library written in Python programming language. This is an additional wrapper for the TensorFlow and Theano frameworks, which allows us to easily switch between them.