Xception-CM: a deep learning approach for non-invasive skin barrier function assessment using dermoscopic images

The integrity of the skin barrier function is critical in the early detection and intervention of various dermatological conditions. Transepidermal water loss (TEWL) is widely regarded as the gold standard for assessing this function. Nevertheless, its measurement fidelity is compromised by pronounced sensitivity to ambient humidity or temperature fluctuations and rigorous protocol standardization requirements. These limitations hinder its scalability and impede continuous real-time monitoring applications in clinical scenarios. To address these limitations, we design a deep neural network for TEWL value prediction based on dermoscopic images. Specifically, we modified the Xception architecture by incorporating a coordinate attention mechanism to enhance spatial feature localization, while employing a multilayer perceptron module to strengthen nonlinear modeling. These components synergistically improve the network's representation power for capturing complex data patterns. To evaluate the proposed model, paired dermoscopic images and TEWL measurements were acquired from standardized left cheek regions of healthy volunteers under controlled experimental conditions. Experimental results demonstrate that the proposed model achieves superior performance in TEWL value prediction, outperforming the state-of-the-art neural network models. Our proposal offers an efficient and non-invasive solution for the clinical evaluation of skin barrier integrity.