Comparing Images Using the Hausdorff Distance
IEEE Transactions on Pattern Analysis and Machine Intelligence
An introduction to ROC analysis
Pattern Recognition Letters - Special issue: ROC analysis in pattern recognition
A narrow band graph partitioning method for skin lesion segmentation
Pattern Recognition
Pattern analysis of dermoscopic images based on Markov random fields
Pattern Recognition
Automatic joint classification and segmentation of whole cell 3D images
Pattern Recognition
Analysis of Pigmented Skin Lesion Border Irregularity Using the Harmonic Wavelet Transform
IMVIP '09 Proceedings of the 2009 13th International Machine Vision and Image Processing Conference
IEEE Transactions on Image Processing
Generalizing Common Tasks in Automated Skin Lesion Diagnosis
IEEE Transactions on Information Technology in Biomedicine
Hi-index | 0.01 |
The boundary irregularity of skin lesions is of clinical significance for the early detection of malignant melanomas and to distinguish them from other lesions such as benign moles. The structural components of the contour are of particular importance. To extract the structure from the contour, wavelet decomposition was used as these components tend to locate in the lower frequency sub-bands. Lesion contours were modeled as signatures with scale normalization to give position and frequency resolution invariance. Energy distributions among different wavelet sub-bands were then analyzed to extract those with significant levels and differences to enable maximum discrimination. Based on the coefficients in the significant sub-bands, structural components from the original contours were modeled, and a set of statistical and geometric irregularity descriptors researched that were applied at each of the significant sub-bands. The effectiveness of the descriptors was measured using the Hausdorff distance between sets of data from melanoma and mole contours. The best descriptor outputs were input to a back projection neural network to construct a combined classifier system. Experimental results showed that thirteen features from four sub-bands produced the best discrimination between sets of melanomas and moles, and that a small training set of nine melanomas and nine moles was optimum.