SegCaps: An efficient SegCaps network-based skin lesion segmentation in dermoscopic images

摘要

This research aims to improve the efficiency of skin lesion segment locations for the given input image of skin cancer using a combination of recently modified segmentation algorithms. Skin lesion segmentation is still a challenging task in medical image analysis because of the low contrast and high noise produced by dermoscopic imaging. Previous works extracted spatially-oriented information but failed in terms of training. They were based on convolutional neural networks (CNNs), which require extensive training time. Current results show 91% to 93% efficiency in segmentation, but the proposed segmentation capsule network (SegCaps) in this research has improved it up to 98% by adding four pre-processing sequential processes in combinations with SegCaps algorithms. The performance of the proposed SegCaps model was evaluated on two different datasets-ISBI 2017 and PH2 and implemented on the MatlabR2017b software. The chosen metrics were Jaccard co-efficient, dice similarity co-efficient, accuracy, sensitivity, and specificity for validation.