Detection and classification of skin lesions based on deep learning from dermatoscop images

It is known that skin diseases are the fourth common disorder that affects approximately one third of the world's population. In skin diseases, the symptoms are the areas on the skin that are different from the normal skin color and pattern and are called lesions. Skin diseases are diseases that can be treated when diagnosed at an early stage. Therefore, early and accurate diagnosis has a critical role in preserving the patient's life. Today, the process of diagnosing skin diseases is based on the dermatologist's visual examination of the skin and making a decision based on what the dermatologist can see. Diagnosis of skin diseases in this way may vary according to the experience of dermatologists. Automated methods enable early diagnosis and interventions. In the literature, automatic methods are suggested and deep learning-based methods are increasingly used in computer aided diagnostic systems. Therefore, within the scope of this thesis, the classification of skin lesions based on deep neural network architecture has been emphasized, and a method that uses convolutional deep neural network architecture and capsule network together, which provides automatic classification of lesions into seven classes with high accuracy from the dermatoscope image, has been proposed. Five different evaluation criteria, namely, specificity, sensitivity, accuracy, F1-score and Mathew correlation coefficient, were used to evaluate the results, and values of 88.10%, 89.18%, 98.11%, 87.25%, and 86.20% were obtained, respectively. These findings showed that the proposed integrated network architecture was successful in automatic classification of skin lesions.