Eczema Region Recognition in Dermatoscopic Images Using PyramidNet Architecture

Abstract In recent years, deep convolutional neural networks (DCNNs) have demonstrated exceptional potential in image-based tasks, including dermatological image analysis. For the task of eczema region identification, a similar approach can be adapted where deep neural networks, structured with multiple convolutional layers, are employed to process high-dimensional image data. Traditional methods in these networks typically increase the feature map dimensions significantly during downsampling to enhance the ability to capture complex features. However, this work introduces a novel strategy, where instead of focusing on substantial feature map expansion solely at downsampling stages, the feature map dimension is gradually increased throughout the entire network. This progressive increase allows the network to capture a broader set of spatial features from the eczema regions, thus improving the model’s robustness and generalization capabilities across varied datasets. Furthermore, a novel residual unit is proposed, which further enhances the network’s ability to recognize and segment eczema-affected areas with greater precision. Experimental evaluations on clinical datasets demonstrate that this new architecture significantly outperforms traditional residual networks in terms of generalization, making it a promising approach for eczema region identification in medical imaging.