Towards Good Features for Skin Cancer Classification

Abstract

Feature extraction is one of the most significant steps in many applications of medical image analysis. Convolutional Neural Network (CNN) has been successfully providing excellent features for some computer vision tasks. Key features for each class have been extracted very well in CNN due to large number of labeled data with various background. These allow accurate and robust classification with CNN in various environment and background. However, it is challenging to collect large volume of data for medical imaging applications with labels. Moreover, it is also challenging to collect data with various backgrounds so that trained CNN with medical data can be potentially biased. To understand what good features are for skin cancer classification, we investigate image features of melanoma and benign images using VGG-19 that is a CNN model trained with general ImageNet database. Skin cancer data with labels (900 images) were obtained from 2016 IEEE ISBI challenge about Skin Lesion. These images were fed into VGG-19 and each image yielded a vector with 1000 elements that are corresponding to classes in. These feature vectors were clustered into 8 categories using k-means algorithm with 3 different distance metrics: Euclidean distance, correlation, and cityblock distance (l1 norm). K-means algorithm with Euclidean distance yielded concentrated clusters that do not have any differentiation power. K-means with other metrics yielded histograms of clusters that are well-spread. Among them, k-means with cityblock yielded the best consistent images for each category (e.g. images with circular patch or pale orange background). Most of these clusters contain both melanoma and benign images, but one category contains mostly benign images (about 100 images). This implies that noncancer related background information may affect CNN for melanoma detection when this database is used without care.