Evaluating automatic colour equalization to preprocess dermoscopic images for classification using a CNN

Sammanfattning: Skin cancer is one of the most prevalent types of cancer and diagnosing of skin lesions are mostly done by visual inspection by a doctor. Lately, computer- aided diagnosis (CAD) has gained popularity and previous studies have with great results utilized a convolutional neural network (CNN) to classify dermoscopic images of different benign and malignant skin lesions. While other studies using CAD tools have investigated the effects of using preprocessing methods on image data before using them in diagnosis classification. Therefore our thesis aims to investigate if preprocessing dermoscopic images of skin lesions before training a CNN in classifying them will improve the accuracy of classification The investigation was conducted by training a CNN on the multi-class problem of classifying dermoscopic images of four different skin lesions. Melanoma and basal cell carcinoma which are malignant and benign keratosis-like lesions and melanocytic nevi which are benign. The dermoscopic images were preprocessed using automatic colour equalization (ACE). The ACE preprocessing was applied to the entire dataset five times each time with different levels of its slope parameter, the contrast tuner of the algorithm. These five datasets together with a dataset not preprocessed with ACE was used to train a CNN model. After 50 epochs of training, the CNN was evaluated on the accuracy of prediction as well as precision, recall and specificity of the four classes. The result indicates that preprocessing images using ACE did not improve the classification accuracy of skin lesions. Additionally, the result suggests that no class is affected more with ACE preprocessing than the others. To further investigate if preprocessing will improve the accuracy of classification the effects of ACE on a different CNN should be conducted. Additionally, if further investigation on the effects of image preprocessing for skin lesion classification is to be conducted, hair removal could be interesting to look into.