Improving Performance of Psoriasis Lesion Segmentation Using Attention-UNet with EfficientNet Encoder

Abstract

Psoriasis is an inflammatory skin disease caused due to the accelerated growth of epidermal tissues giving rise to thick, red, and scaly patches on the skin. It's a lifelong condition that can only be managed with a correct diagnosis and appropriate treatment. The current method of manual assessment for disease diagnosis is tedious and unquantifiable whereas most of the existing computer-aided methods are feature dependent and are less accurate due to the challenging task of lesion segmentation from an uneven background. To overcome these challenges, we propose a fully automatic UNet-based segmentation technique that leverages the benefit of attention and EfficientNet1l as an encoder network for transfer learning. It contains efficiently connected encoders and attention-guided decoders for psoriasis lesion segmentation. The proposed work is evaluated using the Dice Coefficient (DC) and Jaccard Index (JI). The performance result is found to be improved with 0.9590 DC and 0.9215 JI over the existing state-of-the-art method.