Automatic lesion segmentation using atrous convolutional deep neural networks in dermoscopic skin cancer images

Abstract

Background: Melanoma is the most dangerous and aggressive form among skin cancers, exhibiting a high mortality rate worldwide. Biopsy and histopathological analysis are standard procedures for skin cancer detection and prevention in clinical settings. A significant step in the diagnosis process is the deep understanding of the patterns, size, color, and structure of lesions based on images obtained through dermatoscopes for the infected area. However, the manual segmentation of the lesion region is time-consuming because the lesion evolves and changes its shape over time, making its prediction challenging. Moreover, it is challenging to predict melanoma at the initial stage as it closely resembles other skin cancer types that are not malignant as melanoma; thus, automatic segmentation techniques are required to design a computer-aided system for accurate and timely detection.

Methods: As deep learning approaches have gained significant attention in recent years due to their remarkable performance, therefore, in this work, we proposed a novel design of a convolutional neural network (CNN) framework based on atrous convolutions for automatic lesion segmentation. This architecture is built based on the concept of atrous/dilated convolutions which are effective for semantic segmentation. A deep neural network is designed from scratch employing several building blocks consisting of convolutional, batch normalization, leakyReLU layer, and fine-tuned hyperparameters contributing altogether towards higher performance.

Conclusion: The network was tested on three benchmark datasets provided by International Skin Imaging Collaboration (ISIC), i.e., ISIC 2016, ISIC 2017, and ISIC 2018. The experimental results showed that the proposed network achieved an average Jaccard index of 90.4% on ISIC 2016, 81.8% on ISIC 2017, and 89.1% on ISIC 2018 datasets, respectively which is recorded as higher than the top three winners of the ISIC challenge and other state-of-the-art methods. Also, the model successfully extracts lesions from the whole image in one pass in less time, requiring no pre-processing step. The conclusions yielded that network is accurate in performing lesion segmentation on adopted datasets.

Keywords: CNN; Deep learning; Lesion segmentation; Skin cancer.

Fig. 1

Examples of noise artifacts. a Irregular boundaries, b blood vessels, c hairlines, d color illumination, e bubbles, f low contrast

Fig. 2

DilatedSkinNet architecture. An overall layered structure of the proposed method

Fig. 3

Standard convolution. a Rate = 1, Atrous convolution, b rate = 2, c rate = 4

Fig. 4

Exemplary pairs of the segmentation result using DilatedSkinNet. The first row shows original images, second row are gold images and the last row are segmented images a irregular boundaries, b blood vessels, c hairlines, d color illumination, e bubbles, f low contrast

Fig. 5

Comparison of networks. Accuracy versus test samples on the ISIC a 2016, b 2017, c 2018 test sets

Fig. 6

Comparison of networks. Accuracy versus iterations on the ISIC a 2016, b 2017, c 2018 validation sets

Fig. 7

Comparison of networks. Box plots based on Jaccard index of test sets: a ISIC 2016, b ISIC 2017, c ISIC 2018

Fig. 8

Failure cases. Poor segmentation results using the proposed DilatedSkinNet: a Original images, b expected segmentation mask, c segmented outputs, respectively