Automatic Detection of Critical Dermoscopy Features for Melanoma Diagnosis

Period of Performance: 09/12/2007 - 08/31/2008


Phase 2 SBIR

Recipient Firm

Stoecker &associates
Rolla, MO 65401
Principal Investigator


DESCRIPTION (provided by applicant): Malignant melanoma, with an estimated growth in incidence of about 6% per year for decades, causes considerable loss of life. Yet melanoma can be easily cured if detected early. Digital dermoscopy has shown promise for more accurate detection, particularly at an early stage. Recent conferences have highlighted a general agreement on definition of dermoscopic features and moderate agreement on the most useful structural features. Automatic detection of these specific structures that are critical for early diagnosis and are used in various dermoscopic diagnostic algorithms would be desirable. Yet little work has been published on automatic detection of any specific dermoscopic structures. Although specific colors figure prominently in the definition of the most critical dermoscopic structures, little work has been done on finding the specific regions or region combinations in the color space where colors are located, particularly with reference to the surrounding skin. The work in Phase I and after Phase I successfully segmented the border within 5% of the range of the dermatologists' borders, found several highly accurate dermoscopy features, and brought mean diagnostic accuracy on difficult early lesions to a high level. This proposal seeks to develop a digital dermosocopy system by 1) comparing classifiers 2) testing border accuracy and modifying segmentation if needed 3) developing an algorithm that uses a three-dimensional representation of a probability density function to specify single and paired melanoma colors via cluster methods and fuzzy logic techniques 4) identifying critical structural features including brown globules, abrupt border cutoff, granularity, regression, and pigment asymmetry with high accuracy 5) developing a clinical interface for acquisition of images within the clinic 6) testing the new algorithms in six dermatology clinics including two pigmented lesion clinics with both EpiLight and DermLite II Pro dermoscopy images taken in the clinic. Key features of the research include dermatopathology confirmation of specific structures and the use of relative color analysis. If successful, software will be marketed to the growing number of dermatologists with digital dermoscopy capability. The commercial software package will be ready for marketing as a diagnostic adjunct for digital camera dermoscopy attachments. Malignant melanoma, with an estimated growth in incidence of about 6% per year for decades, causes considerable loss of life. Melanoma can be easily cured if detected early, and this project seeks to develop a digital dermoscopy device that can detect very early melanomas. The project goal is to develop inexpensive melanoma detection software and test it in multiple dermatology clinics.