2309 SKIN CONDITIONS AND CROWD-SOURCED HIGH-LEVEL KNOWLEDGE DATASET FOR BUILDING A COMPUTER AIDED DIAGNOSIS SYSTEM

Worldwide, it is believed that there are between 1000 to 2000 skin conditions of which 20 % are difficult to diagnose. An intelligent computer-aided diagnosing system not only helps patients with no or little access to health services but also can benefit typical general practitioners, who have received mini-mal dermatology training. We have built a challenging dataset containing 2309 images from 44 different skin conditions. As most state-of-the-art visual recognition techniques fail to per-form on this dataset, we provide crowd-sourced high-level knowledge of these images to enable the development of a “human in the loop ” approach. This achieves higher recog-nition rates by combining visual features, and human pro-vided information. The high-level knowledge was obtained through employing 361 “Amazon Mechanical Turk ” workers (non-medical experts), who answered a set of predefined per-ceptual questions that represents the humans ’ understanding of lesions in the images. We have made the images, and the high-level knowledge information publicly available. In this paper, we present the data, methods of using it, and perfor-mances of three baseline techniques. This paper and its asso-ciated data will be very useful to facilitate the development of computer aided medical diagnostic techniques. Index Terms — human in the loop, dermatology dataset 1