Automatic telangiectasia analysis in dermoscopy images using adaptive critic design
Article first published online: 5 DEC 2011
© 2011 John Wiley & Sons A/S
Skin Research and Technology
Volume 18, Issue 4, pages 389–396, November 2012
How to Cite
Cheng, B., Stanley, R. J., Stoecker, W. V. and Hinton, K. (2012), Automatic telangiectasia analysis in dermoscopy images using adaptive critic design. Skin Research and Technology, 18: 389–396. doi: 10.1111/j.1600-0846.2011.00584.x
- Issue published online: 13 OCT 2012
- Article first published online: 5 DEC 2011
- Manuscript Accepted: 9 OCT 2011
- image processing;
- adaptive critic design;
Telangiectasia, tiny skin vessels, are important dermoscopy structures used to discriminate basal cell carcinoma (BCC) from benign skin lesions. This research builds off of previously developed image analysis techniques to identify vessels automatically to discriminate benign lesions from BCCs.
A biologically inspired reinforcement learning approach is investigated in an adaptive critic design framework to apply action-dependent heuristic dynamic programming (ADHDP) for discrimination based on computed features using different skin lesion contrast variations to promote the discrimination process. Lesion discrimination results for ADHDP are compared with multilayer perception backpropagation artificial neural networks.
This study uses a data set of 498 dermoscopy skin lesion images of 263 BCCs and 226 competitive benign images as the input sets. This data set is extended from previous research [Cheng et al., Skin Research and Technology, 2011, 17: 278]. Experimental results yielded a diagnostic accuracy as high as 84.6% using the ADHDP approach, providing an 8.03% improvement over a standard multilayer perception method.
We have chosen BCC detection rather than vessel detection as the endpoint. Although vessel detection is inherently easier, BCC detection has potential direct clinical applications. Small BCCs are detectable early by dermoscopy and potentially detectable by the automated methods described in this research.