抽象的な

Automatic Melanoma Detection Using Multi- Stage Neural Networks

Nikhil Cheerla, Debbie Frazier

Skin cancer accounts for more than half of all cancers detected in USA every year. Melanoma is less common, but more aggressive and hence more dangerous than the other types of skin cancers. Even though there has been extensive research in the past 20 years on automatic melanoma detection from skin lesion images, most of the dermatologists still do not have access to this technology. In this paper, a novel system is proposed. The system uses enhanced image processing to segment the images without manual intervention. From the segmented image, it extracts a comprehensive set of features using new and improved techniques. The features were fed automatically to a multistage neural network classifier which achieved greater than 97% sensitivity and greater than 93% specificity. The trained system was tested with lesion images found online and it was able to achieve similar sensitivity. Finally, a new approach that will simplify the entire diagnosis process is discussed. This approach uses Dermlite® DL1 dermatoscope that can be attached to the iPhone. After taking the lesion image with a dermatoscope attached iPhone, the physician gets the diagnosis with a few simple clicks. This system could have widespread ramifications on melanoma diagnosis. It achieves higher sensitivity than previous research and provides an easy to use iPhone based app to detect melanoma in early stages without the need for biopsy.