Publications

The analysis of medical images for skin cancer detection is rising. A fundamental step in image analysis is segmentation. One of the segmentation techniques is thresholding, which is based on finding the optimal threshold value that partitions the image into multiple classes. Otsu’s method, a known thresholding technique searches iteratively for the optimal threshold. It assumes that an image has a Gaussian distribution, which does not always apply to the data in skin cancer images. Skin cancer images usually have a lognormal distribution. We, therefore, propose a Lognormal variant of the Otsu’s sequential method to find the optimal threshold. The sequential search tries all 255 possible values of the threshold, which is time consuming. We therefore propose an iterative Lognormal method, which we found by computing the derivative of Otsu’s optimization formula using a Lognormal distribution. We applied our method on 20 skin cancer images, and we showed that it yields threshold values and segmented images better than the traditional Otsu’s-Gaussian method. Our proposed method uses on average only 6.5 iterations, which saves time and CPU power, compared to the sequential search, which uses 255 iterations. We conclude that Otsu-Lognormal segmentation is more suitable for skin cancer images.