Unsupervised segmentation of skin lesions

Abstract

During the last decades, the incidence rate of cutaneous malignant melanoma, a type of skin cancer developing from melanocytic skin lesions, has risen to alarmingly high levels. As there is no effective treatment for advanced melanoma, recognizing the lesion at an early stage is crucial for successful treatment. A trained expert dermatologist has an accuracy of around 75 % when diagnosing melanoma, for a general physician the number is much lower. Dermoscopy (dermatoscopy, epiluminescence microscopy (ELM)) has a positive effect the accuracy rate, but only when used by trained personnel. The dermoscope is a device consisting of a magnifying glass and polarized light, making the upper layer of the skin translucent.

The need for computer-aided diagnosis of skin lesions is obvious and urgent. With both digital compact cameras and pocket dermoscopes that meet the technical demands for precise capture of colors and patterns in the upper skin layers, the challenge is to develop fast, precise and robust algorithms for the diagnosis of skin lesions. Any unsupervised diagnosis of skin lesions would necessarily start with unsupervised segmentation of lesion and skin.

This master's thesis proposes an algorithm for unsupervised skin-lesion segmentation and the necessary pre-processing. Starting with a digital dermoscopic image of a lesion surrounded by healthy skin, the pre-processing steps are noise filtering, illumination correction and removal of artifacts. A median filter is used for noise removal, because of its edge-preserving capabilities and computer efficiency. When the dermoscope is put in contact with the patient's skin, the angle between the skin and the magnifying glass impacts on the distribution of the light emitted from the diodes attached to the dermoscope. Scalar multiplication with an illumination correction matrix, individually adapted to each image, facilitates the analysis of the image, especially for skin lesions of light color. Artifacts such as scales printed on the glass of the dermoscope, hairs and felt-pen marks on the patient's skin are all obstacles for correct segmentation. This thesis proposes a new, robust and computer effective algorithm for hair removal, based on morphological operations of binary images.

The segmentation algorithm is based on global thresholding and histogram analysis. Unlike most segmentation algorithms based on histogram analysis, the algorithm proposed in this thesis makes no assumptions on the characterization of the lesion mode. From the truecolor RGB image, the first principal component is used as grayscale image. The algorithm searches for the peak of the skin mode, and the skin mode's left bound. The pixel values belonging to the bins to the left of the bound, are regarded as samples from an underlying distribution and the expected value of this distribution is estimated. The value of the pixels in the bin located at equal distance from the expected value of the lesion mode, and the skin-mode peak is used as threshold value. After global thresholding, post-processing is applied to identify the lesion object. The only parameters in this algorithm are the number of bins in the histogram and the shape of the local minimum regarded as skin-mode bound.

The dermoscopic images have been divided into two independent sets; training set and test set. The training set consists of 68 images, and the test set consists of 156 images. 80 of the images from the test set have been evaluated by expert dermatologists by visual inspection.