mCA-Net: modified comprehensive attention convolutional neural network for skin lesion segmentation

ABSTRACT

Skin is  the first line of defense of the human body. Because the skin is exposed to the outside and suffersvarious aggressions , Skin cancer is the most common cancer. Accurate skin lesions image segmentation is essential for skin disease diagnosis and treatment planning. In order toimprove the segmentation results of the recently proposed comprehensive attention convolutional neural network(CA-Net) for skin lesions image segmentation, In this work, we propose a modified medical image segmentation network—modified comprehensive attention convolutional neural network (mCA-Net) to further improve segmentation performance. In particular, we create a new multi-scale channel attention module—MS-CA, which can display more accurate and relevant feature channels on multiple scales. The experiments showsthat our work greatly improve the average segmentation Dice score, accuracy, mean ASSD and mIoU  andenhance the stability of the segmentation model. Through comprehensiveexperiments on the  ISIC 2018 skin lesiondatasets, it is found that our proposed mCA-Netnetwork compared with CA-Net,improve the average segmentation Dice score from 92.08% to 93.56%, the average accuracy score of skin lesions from 92.68% to 93.32% and the mIoU from 85.32% increased to 87.89%. The segmentation results have been significantly optimized.

KEYWORDS:

• Timage segmentation
• attention
• melanoma

Acknowledgments

This work is supported by Science and Technology department of Xinjiang Uyghur Autonomous Region (2020E0234 and 2021B03001-4).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work is supported by Key R & D project of Xinjiang Uygur Autonomous Region(2021B03001-4) and Science and Technology department of Xinjiang Uyghur Autonomous Region(2020E0234).