•  
  •  
 

Abstract

Skin lesion segmentation is an essential step toward accurate skin lesion diagnosis. The need to automate Skin lesion segmentation on the one hand, and the challenges it faces, on the other hand, have made it a growing area of research and focus. Automation of skin lesion segmentation helps reduce the effort and time needed for diagnosis and treatment and helps make better utilization of available data and shared experiences. The challenges faced by the automation of skin lesion segmentation can be broadly defined by (but not limited to); variations in texture, shape, and size for skin lesions and the low contrast between the lesion and surrounding skin.

The rise of deep learning has significantly improved the semantic segmentation results in medical imaging. U-Net structure with encoder and decoder approach is one of the most successful deep learning models for medical image segmentation. This paper introduces two models based on U-shaped structures: AlexUnet and AlexUnet+.

AlxUnet is a light U-Net model with an encoder based on pre-trained AlexNet on the ImageNet database. It significant-ly reduces memory consumption and the number of parameters, thus reducing the required FLOPS by eight times. In Alexunet+, another encoder was added to the AlxUnet structure that used pre-trained VGG11 on ImageNet. It is al-lowed to aggregate the feature maps obtained from two encoders to be used in the decoder.

AlxUnet and AlxUnet+ models were evaluated using three publicly available databases provided by the International Skin Imaging Collaboration, ISIC 2016, ISIC 2017, and ISIC 2018. Sensitivity, specificity, Jaccard similarity index, and dice similarity were used as performance metrics. Then, obtained structures were compared with U-Net, and many deep learning segmentation networks that were recently built for skin lesion segmentation. AlxUnet outperformed U-Net and produced acceptable results compared with the other networks. AlexUnet+ produced a more robust result and outperformed other networks.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Share

COinS