Optimum blend of fractal methods for automatic malignancy determination in dermoscopy images

The most important step is early and effective diagnosis in the treatment of melanoma which constitutes the biggest part of skin cancers results in death. At the present time, dermatologists take dermoscopy images, visually examine these images and draw the lesion borders manually for a future reference. With the naked-eye, it is not easy to recognize compartments and tissue structures. Additionally, this procedure is also tedious, biased, and failure rate is high. Moreover, even with the help of dermoscopy, 70% of melanoma claims are still a false-negative diagnosis. This is the motivation for computer assisted diagnosis (CAD) techniques to help dermatologists to reduce possible unlikelihood, to standardize the results and also to speed up the process. The techniques which are developed by Fractal Methods determine irregularities on the lesion borders. Our aim is to designate which Fractal Methods are more effective on determining malignant lesions to minimize false-negative, false-positive and total-false diagnosis. For this purpose, we develop four different mixed integer programming (MIP) classification models, and then applied these models on the dataset of 100 patients. First, we determine the optimum usage rate of Fractal Methods for each classification model using randomly selected 50 patients (training group). Later, true diagnosis performance of each classification model is evaluated using the remaining 50 patients (testing group) and the optimum usage rate of Fractal Methods which is already found with the "training group". It is observed that the optimum usage rate of fractal methods gives 80% success rate in the best scenario but the average on the repeated tests with these optimum usage rate of fractal methods that we selected is 73.4%.