Automated segmentation of skin lesions as of digitally recorded images is a crucial procedure to diagnose skin diseases accurately. This paper proposes a segmentation model for skin lesions centered on Deep Convolutional Neural Network (DCNN) for melanoma, squamous, basal, keratosis, dermatofibroma, and vascular types of skin diseases. The DCNN is trained from scratch instead of pre-trained networks with different layers among variations in pooling and activation functions. The comparison of the proposed model is made with the winner of the ISIC 2018 challenge task1(skin lesion segmentation) and other methods. The experiments are performed on challenge datasets and shown better segmentation results. The main contribution is developing an automated segmentation model, evaluating performance, and comparing it with other state-of-art methods. The essence of the proposed work is the simple network architecture and its excellent results. It outperforms by obtaining a Jaccard index of 87%, dice similarity coefficient of 91%, the accuracy of 94%, recall of 94% and precision of 89%.
Paper
Full text
An Effective Skin Disease Segmentation Model Based on Deep Convolutional Neural Network
Semantic Scholar · Medicine · 2022
Abstract
Automated segmentation of skin lesions as of digitally recorded images is a crucial procedure to diagnose skin diseases accurately. This paper proposes a segmentation model for skin lesions centered on Deep Convolutional Neural Network (DCNN) for melanoma, squamous, basal, keratosis, dermatofibroma, and vascular types of skin diseases. The DCNN is trained from scratch instead of pre-trained networks with different layers among variations in pooling and activation functions. The comparison of the proposed model is made with the winner of the ISIC 2018 challenge task1(skin lesion segmentation) and other methods. The experiments are performed on challenge datasets and shown better segmentation results. The main contribution is developing an automated segmentation model, evaluating performance, and comparing it with other state-of-art methods. The essence of the proposed work is the simple network architecture and its excellent results. It outperforms by obtaining a Jaccard index of 87%, dice similarity coefficient of 91%, the accuracy of 94%, recall of 94% and precision of 89%.