Medical image learning from a few/few training samples: Melanoma segmentation study

Melanoma is a type of life-threatening skin cancer which starts in melanocytes. The five-year survival rate of melanoma will increase to 98 percent in case of being detected and diagnosed before involving lymph nodes. Hence, early detection of melanoma has a pivotal role in decreasing the related fatal rate. One of the most popular classes of early detection methods is dermoscopy image analysis, where color photos of skin lesions are processed to determine if a mole is malignant or not. In order to properly implement these techniques, these dermoscopy images must be separated into foreground (the mole) and background (the surrounding skin) in a process known as segmentation. Multiple Random Walker (MRW) and deep learning approaches have been used for melanoma segmentation in this paper. We design a MRW-based system, a semi-automatic approach, for segmentation of dermoscopy images. We investigate the effect of number of walkers on the results and find the optimal gradient calculation algorithm for our setup. We have also applied three state-of-the-art Convolutional Neural Networks (CNNs) including SegNet, U-Net and DeconvNet, designed for segmentation, and analyzed the procedure of segmentation and their performance on delineating melanoma. Finally, we investigate the selection of the best approach adaptively based on the size of the available training masked images. The results reveal that MRW-based segmentation approach is a promising selection when training set images are limited while the CNN architectures are decent choice in the presence of large training set.

Duke Scholars

Author Ming-Chun Huang DKU Faculty