Development of computational image processing algorithms for detecting morphological features of melanoma

This thesis research is focused on advancing image processing techniques and algorithms used for detecting skin melanoma. We have modified previous image processing approaches and developed computational algorithms for quantifying morphological features of a mole image. Applying the algorithms to 20 mole images downloaded from educational websites, we have identified three cut-off ratios to distinguish melanoma images from benign mole images. More specifically, the higher the boundary irregularity ratio, and/or the asymmetry ratio, and/or the color variation ratio, the high chance the mole is melanoma leading. The irregularity ratio cutoff is identified as 1.96, suggesting 96% more circumference length than that of a circle with the same area. One finds that the cut-off ratio for assessing asymmetry of the mole image is 0.109, representing the degree of asymmetry as approximately 11% to place a mole image into the melanoma group. Evaluation of the color variation of the moles leads to a cut-off ratio of the color variation as 0.334. Statistical analyses have been performed to determine the confidence of cut-off ratios, varying from 63% to 81%, for placing a mole image into its correct groups. The algorithms have also been implemented to assess "changes" of mole images over time observed by a dermatologist. Using a +/-14% as the definition of changes, the algorithm identifies 9 of the 10 mole images as changed over time. Among the irregularity, asymmetry, color variation, and size ratios, 5 out of the 9 moles have shown changes in one ratio, 2 out the 9 moles have experience changes in two ratios, 1 mole has shown changes in three ratios, and only 1 mole shows changes in all four ratios. The computational results are consistent with the general observations that human eyes are sensitive to size changes and color variation changes, and may not be very good to distinguish changes in border irregularity and asymmetry. The developed algorithms can be helpful to assist a physician in evaluating subtle changes of mole images that may not be very sensitive to the eyes.