Improvement proposals for lesion segmentation on dermoscopic images

Abstract

Purpose: The aim of this study is to present new proposals to improve the task of lesion segmentation on dermoscopic images. These proposals are aimed at the model that performs the segmentation and the data augmentationmethod that supports the model within the segmentation process. Theory and Methods: The implemented segmentation architecture (EnecaU-Net) has two encoder paths and one decoder section. The model we propose receives the same input samples at the same time through two encoder paths, the up-sampling process is used in the first encoder part of the model (in order to obtain a precise segmentation ofthe region of interest). In addition, the Efficient Channel Attention (ECA) module has also been added to the second encoder section to refine the dense information in the image and process useful information. The data augmentation method applied in this study is based on the principle of adding (combining, mixing) randomly selected samples from the International Skin Imaging Collaboration (ISIC-2016) training images according to the value lambda=0.4. The obtained 2000 images were trained at 50 epochs with a size of 128 x 128 at the gray level. No pre-processing techniques were applied to the images before the training andtesting processes in order to remove the factors that would complicate the segmentation process, such as skin hairs or air/oil bubbles, from the images. Results: Experimental evaluations' results show that the proposed model used together with the mix data augmentation method achieves high results after training and then for testing. The proposed model achieves training validation results with an accuracy of 97.36%, a sensitivity of 94.64%, a specificity of 98.16%, a Dice similarity coefficient of 94.11%, and a Jaccard index of 88.89% on the ISIC-2016 dataset; and the proposed model achieves test results with an accuracy of 94.19%, a sensitivity of 90.73%, a specificity of94.78%, a Dice similarity coefficient of 88.05%, and a Jaccard index of 80.30% on the ISIC-2016 dataset. Conclusion: The segmentation success of the EnecaU-Net model applied for lesion segmentation on dermoscopic images is high, and the applied mix data augmentation method improves the segmentation success of the EnecaU-Net model.