Skin Lesion Segmentation from Dermoscopic Images Using Convolutional Neural Network

Abstract

Clinical treatment of skin lesion is primarily dependent on timely detection and delimitation of lesion boundaries for accurate cancerous region localization. Prevalence of skin cancer is on the higher side, especially that of melanoma, which is aggressive in nature due to its high metastasis rate. Therefore, timely diagnosis is critical for its treatment before the onset of malignancy. To address this problem, medical imaging is used for the analysis and segmentation of lesion boundaries from dermoscopic images. Various methods have been used, ranging from visual inspection to the textural analysis of the images. However, accuracy of these methods is low for proper clinical treatment because of the sensitivity involved in surgical procedures or drug application. This presents an opportunity to develop an automated model with good accuracy so that it may be used in a clinical setting. This paper proposes an automated method for segmenting lesion boundaries that combines two architectures, the U-Net and the ResNet, collectively called Res-Unet. Moreover, we also used image inpainting for hair removal, which improved the segmentation results significantly. We trained our model on the ISIC 2017 dataset and validated it on the ISIC 2017 test set as well as the PH² dataset. Our proposed model attained a Jaccard Index of 0.772 on the ISIC 2017 test set and 0.854 on the PH² dataset, which are comparable results to the current available state-of-the-art techniques.

Keywords: Jaccard Index; ROC curve; ResNet; U-Net; convolutional neural networks; dermoscopic images; image inpainting; melanoma.

Figure 1

Examples of the ISIC-17 dataset (a) and PH² dataset (b).

Figure 2

Representation of a single image as it is passed through the hair removal algorithm. Left to right: (a) Input image; (b) Grayscale image; (c) Cross shaped structuring element employed during morphological operations; (d) Image obtained after applying black top-hat filter; (e) Image after thresholding; (f). Final image obtained as output.

Figure 3

Schematic diagram representing UResNet-50. The encoder shown on the left is the ResNet-50, while the U-Net decoder is shown on the right. Given in the parenthesis is the channel dimensions of the incoming feature maps to each block. Arrows are defined in the legend.

Figure 4

Training and validation accuracy of the proposed convolutional neural network model for 70 epochs.

Figure 5

Example results of multiple patients. The first row contains the original images of five patients from the test set. The second row contains corresponding ground truths as provided. The third row contains predicted masks from the proposed method.

Figure 6

Receiver operative characteristics (ROC) curve generated on the ISIC-17 test set.