ABSTRACT The accurate skin cancer segmentation is essential for clinical diagnosis, especially when it comes to dealing with ambiguous and low‐contrast target regions. Moreover, the fuzzy borders, irregular and noisy shapes also become a hurdle in the task of segmentation. To address these challenges, an encoder–decoder architecture is proposed that integrates Squeeze‐and‐Excitation (SE) blocks in the encoder phase and similarly in the decoder that also integrates SE for channel attention as well as the Attention Gates (AG) for spatial attention. Every block of the encoder extracts the features using convolutional layers and adjusts them through SE modules, which enhance the informative channels. The bottleneck refines the global contextual features through the same strategy. During the process of decoding, AGs filter the encoder attributes based on the context of the decoder, which allows only the relevant spatial information to flow through. The decoder part also utilizes the SE blocks for the enhancement of the reconstruction quality. The final output map is generated by the convolution and sigmoid activation. The method is evaluated on the six publicly available datasets, including PH2, HAM10000, ISIC2018, ISIC2019, ISIC2020, and MILK10K. The proposed architecture provides better outcomes as compared to several existing methods.
Zafar et al. (Fri,) studied this question.