An Efficient Image Processing and Machine Learning based Technique for Skin Lesion Segmentation and Classification

Skin cancer is one of the most lethal types of cancer that has grown extensively over the last decade and can lead to death if not treated on time. However, if skin cancer is diagnosed and treated at an early stage, it can be curable. Several image processing and machine learning methods have been considered to detect and classify skin cancer lesions (melanoma) accurately. However, a lower contrast of images also affects the segmentation efficiency and further increments classification error. Thus, in this work, a simple yet effective image processing and machine learning based technique has been proposed for skin lesion segmentation and classification to overcome this problem. The proposed technique increases the segmentation accuracy in its pre-processing stage, removing noise and hair and enhance image contrast by adjusting intensity values of RGB channels. Otsu thresholding and image subtraction methods are applied to extract the region of interest and segmented lesion area. Morphological operations are performed to remove noisy pixels and reshape the segmented image at the post-processing stage. For extraction of image features, color and ABCD features are applied. The PH2 dataset is used in this work, consisting of imbalanced classes; therefore, Synthetic Minority Over-sampling Technique (SMOTE) is used to balance the distribution of the dataset. A supervised machine learning classifier further uses the extracted image features to classify skin lesions. The proposed technique accurately segments the lesion with an accuracy of 90.25% and classifies them into melanoma and non-melanoma with an average accuracy of 98.13% using the Adaboost classifier.