Unsupervised skin lesion classification and matching

Show full item record

Title: Unsupervised skin lesion classification and matching
Author: Yandow-Reilly, Paula
Abstract: According to the American Cancer Society (ACS), since 1973, the mortality rate for melanoma has increased by 44%. The number of serious skin cancers diagnosed has also more than doubled in that same period. Even though serious skin cancers (melanoma) account for only 4% of skin cancer diagnoses (and skin cancer is the most common cancer) it is responsible for almost all (79%) cancer deaths. The ACS reports about 7,300 people in the United States are expected to die of melanomas in 2002, other sources put the number as high as 7,800. There are about 130,000 cases of melanoma worldwide, and about 37,000 related deaths. Many physicians think the increase in melanoma diagnoses represents an epidemic. Currently, there is work to improve diagnostics once a lesion comes under suspicion, and there are also systems to do whole body images of skin lesions. Where there seems to be a gap is in tracking and classifying the lesions in image histories. The critical problem is not so much how to treat the lesion once its discovered, but to detect it in the first place. In addition, in the classification systems encountered, there didn't seem to be any using all combinations of color, texture, and shape, any or all of which can help detect a malignant growth. Since almost all lesions are slow-growing, and very often on the back, it can be difficult for both patient and doctor to detect when a lesion has begun to change, which is one of the first warning signs of skin cancer. This work is comprised of an analysis system written in Matlab, which pre-processes the image, removing background artifacts via morphological operations to segment the lesion. The lesion is then processed for shape, color content, and texture. This occurred for a small database of images comprising melanomas, dysplastic nevi, and moles, and 10 feature vectors were captured for each image along with the filename and matching diagnosis. Additional images were procured from the web, and also from photographs of individuals using a Cannon EOS Rebel G, which were scanned in using an Acer ScanPrisa 640U. These images were then processed with the same software used for the database images. The results were classified based on these feature vectors and assigned a FWL (Feature Warning Level). Lastly, the input results were compared to the database for matches within a range for similarity. The closest match (if within a reasonable range) is reported. This system could be attached to existing tracking systems (like MoleMap) or used as a stand alone tracking tool for dermatologists. Any change in one of the feature vectors, or in a group of features could trigger a closer look by the physician. According to literature, and a dialog with a dermatologist, history is the one of the most critical factors in early detection, when the cancer can be completely cured.
Record URI: http://hdl.handle.net/1850/15214
Date: 2003

Files in this item

Files Size Format View
PYandow-ReillyThesis02-2003.pdf 18.44Mb PDF View/Open

The following license files are associated with this item:

This item appears in the following Collection(s)

Show full item record

Search RIT DML


Advanced Search

Browse