Article Preview
TopIntroduction
The recent tremendous growth in computer technology has brought a substantial increase in storage needs of digital images. In the medical field, millions of images are generated by clinicians every year. The storage of such medical image data (dental, endoscopy, dermoscopy, skull, MRI, CT, ultrasound, x-ray, mammography, cytological specimens) is relatively straightforward. However, accessing and searching medical image databases through traditional text-based image retrieval techniques is intrinsically harder. The most common systems used for text-based medical image retrieval refer to Foundational Model of Anatomy (FMA) (Rosse & Mejino, 2003), Radiology Lexicon (RadLex) (Langlotz, 2007), DermQuest resource (available at http://www.dermquest.com) and so on. These systems present several drawbacks such as tedious and time consuming image annotation. Moreover, the manual description of the medical images is unscalable and subjective. This is due to the nature of human perception. Thus, text-based image retrieval techniques lead to inaccuracy in the skin lesion image retrieval process. Therefore, Content-Based Medical Image Retrieval (CBMIR) is an interesting alternative.
CBMIR opens the way for clinicians to retrieve from prior known cases the images that contain regions with features similar to the query. With the knowledge of disease entities that match with the features of the query image and associated diagnostic information, the clinicians can come up with a diagnostic decision. So far, a variety of CBMIR systems have been developed: IRMA (Thies, Güld, Fischer & Lehmann, 2005), BRISC (Lam, Disney, Raicu, Furst & Channin, 2007), Cervigram Finder (Xue, Long, Antani, Jeronimo & Thoma, 2008), to name a few.
These systems are mostly prototypes used for lab experiments rather than in clinical practice. Furthermore, their performance is still limited, mainly because of semantic gap, which expresses the inconsistency between abstract visual feature representations of skin lesion images and their actual meaning described by clinicians.
Regarding the limits affecting CBIR systems, Relevance Feedback (RF) has been introduced in the retrieval process in order to enhance the results. It is viewed as an intelligent way of bridging the semantic gap and reflecting the user’s need.
The fundamental goal of RF is to establish an interaction between the user and the retrieval system to iteratively refine retrieval results, on the basis of the user assessment of the results.
A comprehensive survey of CBIR techniques using RF methods can be found in (Datta, Joshi, Li & Wang, 2008).
In this paper, we propose an intelligent Content-Based Dermoscopic Image Retrieval system with Relevance Feedback (CBDIR with RF) for melanoma diagnosis. This is accomplished using Linear Kernel based active Support Vector Machines (LKSVMactive) algorithm. Intuitively, LKSVMactive works by combining three ideas.
- •
First, an SVM that captures the image query concept by separating relevant images from irrelevant ones with a hyperplane in a projected space. The projected points on one side of the hyperplane are considered relevant to the query concept, while the remaining point are irrelevant;
- •
Second, an active SVM that learns quickly the classifier via active learning sampling. It selects the closest images to the SVM hyperplane to enhance the SVM classifier. This step ensures fast convergence toward the query concept in a small number of feedback rounds;
- •
Third, a histogram intersection similarity measure to derive the most relevant images with respect to the query image.
The proposed system is characterized by the following aspects.