Automated skin lesion segmentation using multi-scale feature extraction scheme and dual-attention mechanism

TL;DR

This paper introduces a deep learning model with multi-scale feature extraction and dual-attention mechanisms for automatic skin lesion segmentation, effectively handling challenges like poor contrast and unclear boundaries.

Contribution

The work presents a novel deep learning architecture combining multi-scale features and dual-attention modules specifically designed for skin lesion segmentation.

Findings

Outperformed existing models on ISIC2018 and ISBI2017 datasets.

Achieved top rankings in two skin lesion segmentation competitions.

Demonstrated robustness against image artifacts and boundary ambiguities.

Abstract

Segmenting skin lesions from dermoscopic images is essential for diagnosing skin cancer. But the automatic segmentation of these lesions is complicated due to the poor contrast between the background and the lesion, image artifacts, and unclear lesion boundaries. In this work, we present a deep learning model for the segmentation of skin lesions from dermoscopic images. To deal with the challenges of skin lesion characteristics, we designed a multi-scale feature extraction module for extracting the discriminative features. Further in this work, two attention mechanisms are developed to refine the post-upsampled features and the features extracted by the encoder. This model is evaluated using the ISIC2018 and ISBI2017 datasets. The proposed model outperformed all the existing works and the top-ranked models in two competitions.