Classification of dermoscopic images using neural networks

Melanoma is amongst one of the deadliest form of cancer. American Cancer Society reports that in 2016, about 76,380 new melanomas will be diagnosed and about 10,130 people are expected to die of melanoma in U.S. Even though melanoma constitutes less than 2% of skin cancer, this type of skin cancer represents the majority of deaths caused by skin cancer. The most basic way to reduce mortality due to melanoma is early diagnosis. Dermatologists, use dermoscopy for scanning the skin and as a result of this scan melanomas can be detected. A gel is applied on the lesion, and depending on the lesion enlargement devices, the lesion is magnified 6-100 times. This magnified view is used to evaluate based on various medical diagnostic procedures, such as ABCD rule, seven points check-list, and Menzies method. ABCD rule allows assigning a score for each lesion. This score is calculated using four different characteristics (Asymmetry, Border, Colors, Differential Structures) of the lesion. Menzies method includes two different features: negative (symmetrical patterns, color) and positive (blue-white veil, atypical structures, etc.). The presence of positive features indicate melanoma. Finally, the lesions are scored by seven-point check-list. However, in this method, only, lesions are checked for whether there is anyone of different structures (atypical pigment network, irregular streaks, etc.) and assigned scores 2 or 1 for each structure. If the total score is higher than 3 this may indicate that the lesion is melanoma. These methods are still applied subjectively. Dermoscopy maximized rate of identification of the disease, however usually diagnosis of the disease is based on physician's experience. Since the determination is based on the human vision and experience, there is huge effort to determine of the disease using digital techniques. Dermascopic image analysis with computers can be used to prevent subjective decisions and to maximize the accuracy of melanoma detection. In this thesis, new methods that are based on ABCD rules-motivated features, statistical features and textural features and used to classify dermoscopic lesion images are proposed and compared. Dermoscopic lesion images are classified using different features. ABCD Rules-motivated features contain radius histogram of lesions, relational size of lesions and the magnitude and the phase of Fourier descriptors of the border of segmented lesions. All of these features are used to classify the lesion images together and separately. Statistical features of texture contain first-order statistical features and second-order statistical features of the lesion images. The lesion images are divided 128x128 patches and first-order and second-order features are extracted from these patches. First-order statistical features contain mean, variance, skewness, kurtosis, energy, and entropy. Second-order features contain correlation, contrast, inverse difference, entropy, and maximum probability. These statistical features are used to classify lesion images. To extract textural features, we use wavelet transformation. In wavelet transformation features contains the percentages of energy corresponding to the horizontal, vertical, and diagonal details. These energy percentages are used as dermoscopic image features. PH2 dermoscopic images are used as database. This database contains 200 images in total. 40 of them are melanoma, 160 of them are not melanoma. 180 of the mareused to train ANN with 10-fold cross validation. 20 of them are used as a test set. According to our result the best method to classify our dermoscopic image database is ABCD rules-motivated methods. Using ABCD Rules-motivated methods, sensitivity=72.22% and specificity=93.75% are accomplished.