Automated Melanoma Detection in Dermoscopic Images

Cancer, with its varying and hard to detect types, became one of the most dangerous diseases for humans. Melanoma is a type of skin cancer that has the most mortality rate among its type. The usual melanoma detection process is based on awareness of the patient and the experience of the visual investigator. Even though the invention of dermoscopes reduce its effects, ?subjectivity? problem plays a huge role on the detection accuracy, which creates a need for automated detection. In this thesis, history of automated melanoma detection on dermoscopic images and caveats of present frameworks are studied. Different approaches to overcome these caveats are explored. As a result, a new melanoma detection algorithm based on Bag of Visual Words (BoVW) concept, which combines traditional methods with new age deep learning techniques, is created. The performance of the new algorithm is tested on the popular International Skin Imaging Collaboration (ISIC) Challenge 2017 dataset, which yielded tremendously good results. With 96.2% accuracy and more importantly with 99.8% sensitivity, it surpassed all other entries in the ISIC 2017 Leaderboard. Since, sensitivity represents the algorithm?s success on correctly classifying melanoma cases, this success places the algorithm on a special place in the domain. Lastly, future directions on the domain are explored on the terms of increasing the performance of the newly born algorithm further.