Hybrid Pre-Trained Models with Machine Learning Classifier for Skin Cancer Classification

Classification of skin lesions can be challenging because of the subtle variation on the skin surface due to the appearance of these lesions. In the other hand, the digital dermoscopy has been widely used by dermatologist to diagnose cancer. For accurate detection, the clinicians should have a lot of experiences, but human nature is prone to error, forgetfulness, tension, and speed in diagnosis, all of which can affect the accuracy of detection. For these reasons, automated skin lesion classification has been used by many researchers to help dermatologists in making the right and accurate decision during diagnosis. A preprocessing image pipeline has been applied in this work prior to classification, including image enhancement, image normalization, and resizing. By established transfer learning on ImageNet weights of pre-trained and fine-tuning the model to meet our purpose. Five pre-trained models of CNN are used with four type of machine learning classifiers on HAM10000 dataset. This paper suggests a hybrid model for feature extraction, which then feeds these features to the classifier to classify dermoscopy images as either benign or malignant. The highest reported score by concatenated DenseNet201 and Mobil Net with SVM classifier are 87.8%, 86.956%, 87.912%, 87.755%, 87.43%, 100%, 94%, and 90% in term of Precision, Sensitivity, Specificity, F1-score, AUC for training data, AUC for validation data, and AUC for testing data, respectively.