|
Melanoma is a malignant tumor of melanocytes. Though less common than other types of skin cancer, it is considered the deadliest form not only because it causes the majority (75%) of deaths related to skin cancer, but also because it can metastasize to other organs in the body. The lifetime risk of developing a melanoma has been skyrocketing in the United States, growing from 1:1,500 people in 1953 to nearly 1:100 in 1996. In fact, the American Cancer Society (ACS) reported that in 2010 an estimated 68,130 new cases of melanoma were diagnosed, resulting in a total of 8,700 deaths. The National Cancer Institute (NCI) at the National Institutes of Health (NIH) estimates that in 2012 the new cases of confirmed diagnosis and resulting deaths from melanoma are 76,250 and 9,180, respectively. On the other hand, the survival rate of melanoma is above 85% if detected no later than stage I/II, and the cost for an early-stage melanoma treatment is as low as about $1,800 per patient. Advances in imaging technologies enabled non-invasive skin cancer screening and early detection using pigmented skin lesion (i.e., moles) images. Furthermore, such non-invasive and low cost imaging systems (such as dermoscopy) made it possible for mole screening in the public-health setting by general doctors. However, even the most experienced dermatologist is challenged to perform an assessment and give the correct diagnosis using these images. Because moles vary in size, color, shape, and texture pattern, and the diagnosis severity ranges from completely benign to aggressive and lethal (not to mention the sheer number of lesions that need to be evaluated), such screening is still nonexistent in the current healthcare system even though reports recognize the paramount importance of screening and early detection. An automated melanoma screening and early detection system uses computer-generated quantitative features from dermoscopic skin lesion images to detect the feature models (i.e., signature) of various types of skin cancer.
This chapter begins with an overview of skin-lesion-imaging-based diagnosis systems as well as the available dermoscopy skin lesion datasets in Section 7.1. We present the architecture of an automated melanoma screening and early detection system, AutoScan, and discuss the function of each component in Section 7.2. Section 7.3 describes typical computergenerated feature sets used in melanoma detection systems before presenting our effort to incorporate dermatologists' domain knowledge as a high-level feature. Section 7.4 presents our method to integrate feature selection and decision making by selecting the optimum feature set to build an "on-the-fly" feature model based on their holistic predictive performance. Section 7.5 concludes the chapter and presents future research directions.
|
