Lesion Segmentation and Automated Melanoma Detection using Deep Convolutional Neural Networks and XGBoost

Melanoma is a lethal form of skin cancer. It is one of the leading causes of deaths related to skin cancer. Most clinical studies show that early diagnosis of melanoma can improve patient outcomes. Currently, the most effective mean of analyzing lesions involves the use of dermoscopic imaging. However, this process is highly Subjective as it is dependent on the proficiency levels and experience of the specialist. Our research aims at minimizing the level of uncertainty and Subjectivity in human assessment. We propose a deep learning approach to model lesion patterns with the goal of performing automated melanoma detection and lesion segmentation from skin images. We use an ensemble of deep learning models to combine multiple hypothesis into a single decision point. This mimics the real world approach where the specialist will typically consult other specialists to cross reference and double check their diagnosis before consulting with the patient. We built different deep learning models using the same dataset with extensive data augmentation. For melanoma detection, deep convolutional neural networks including Inception-v4, ResNet-152 and DenseNet-161 were trained for melanoma classification and seborrheic keratosis classification. For lesion segmentation, U-Net and U-Net with VGG-16 Encoder were trained to produce segmentation masks. The proposed method was evaluated on the ISIC 2017 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset, and our model ranked 5
^th
in classification and 8
^th
in segmentation among 23 and 21 international teams, respectively.