How Users Perceive Content-Based Image Retrieval for Identifying Skin Images

摘要

Content-Based Image Retrieval (CBIR) is an application of computer vision techniques for searching an existing database for visually similar entries to a specific query image. One application of CBIR in the dermatology domain is displaying a set of visually similar images with a pathology-confirmed diagnosis for a given query skin image. Recently, CBIR algorithms using machine learning with high accuracy rates have gained more attention since researchers have reported they have the potential to help physicians, patients, and other users make trustworthy and accurate classifications of skin diseases based on visually similar cases. However, we do not have many insights into how interactive CBIR tools are actually perceived by end users. We present the design and evaluation of a CBIR user interface and investigate users' classification accuracy on dermoscopy images and explore users' perception of confidence and trust. Our study with 16 novice users for a given set of annotated dermoscopy images indicates that, in general, CBIR enables novices to make a significantly more accurate classification on a new skin lesion image from four commonly-observed categories: Nevus, Seborrheic Keratosis, Basal Cell Carcinoma, and Malignant Melanoma.