Combination of 3D skin surface texture features and 2D ABCD features for improved melanoma diagnosis

Abstract

Two-dimensional asymmetry, border irregularity, colour variegation and diameter (ABCD) features are important indicators currently used for computer-assisted diagnosis of malignant melanoma (MM); however, they often prove to be insufficient to make a convincing diagnosis. Previous work has demonstrated that 3D skin surface normal features in the form of tilt and slant pattern disruptions are promising new features independent from the existing 2D ABCD features. This work investigates that whether improved lesion classification can be achieved by combining the 3D features with the 2D ABCD features. Experiments using a nonlinear support vector machine classifier show that many combinations of the 2D ABCD features and the 3D features can give substantially better classification accuracy than using (1) single features and (2) many combinations of the 2D ABCD features. The best 2D and 3D feature combination includes the overall 3D skin surface disruption, the asymmetry and all the three colour channel features. It gives an overall 87.8 % successful classification, which is better than the best single feature with 78.0 % and the best 2D feature combination with 83.1 %. These demonstrate that (1) the 3D features have additive values to improve the existing lesion classification and (2) combining the 3D feature with all the 2D features does not lead to the best lesion classification. The two ABCD features not selected by the best 2D and 3D combination, namely (1) the border feature and (2) the diameter feature, were also studied in separate experiments. It found that inclusion of either feature in the 2D and 3D combination can successfully classify 3 out of 4 lesion groups. The only one group not accurately classified by either feature can be classified satisfactorily by the other. In both cases, they have shown better classification performances than those without the 3D feature in the combinations. This further demonstrates that (1) the 3D feature can be used to improve the existing 2D-based diagnosis and (2) including the 3D feature with subsets of the 2D features can be used in distinguishing different benign lesion classes from MM. It is envisaged that classification performance may be further improved if different 2D and 3D feature subsets demonstrated in this study are used in different stages to target different benign lesion classes in future studies.

Appendix 1: Locating a lesion’s centre of mass and principal axis using moment

The moment of order (p + q) for an M × N digital image is given by

$${m_{pq} = \sum\limits_{m = 1}^{M} {\sum\limits_{n = 1}^{N} {m^{p} n^{q} S(m,n)} } }$$

(20)

The centralised moments are given by

$${m_{pq}^{c} = \sum\limits_{m = 1}^{M} {\sum\limits_{n = 1}^{N} {(m - m_{c} )^{p} (n - n_{c} )^{q} S(m,n)} } }$$

(21)

where (m _c, n _c) is the centre of the mass, which is defined as

$${m_{c} = \frac{{m_{10} }}{{m_{00} }}}\;\;{\text{and}}\;\;{n_{c} = \frac{{m_{01} }}{{m_{00} }}}$$

(22)

and S(m, n) is a binary image generated as

$${S(m,n) = \left\{ {\begin{array}{*{20}l} 1 \hfill &\quad {{\text{if}}\;(m,n) \in S_{\text{l}} } \hfill \\ 0 \hfill &\quad {\text{otherwise}} \hfill \\ \end{array} } \right.}$$

(23)

where S _l denotes the lesion region. Then, the direction of the principle axis of a lesion is given by

$${\omega_{0} = \frac{1}{2}\tan^{ - 1} \left( {\frac{{2m_{11}^{c} }}{{m_{20}^{c} - m_{02}^{c} }}} \right)}$$

(24)