An inversion approach for non-invasive detection of subcutaneous structure and temperature based on 1D residual neural network

Laser skin dermatology has attracted increasing attention with awareness of the aesthetic sense. Real-time detection of subcutaneous vessel temperature and structure during laser treatment is essential for efficacy and safety of surgery. By using 1D residual neural network to improve the accuracy of feature extraction, an inverse method model with outstanding reconstruction performance and noise insensitivity was proposed to estimate vascular parameters including temperature, depth and diameter of blood vessel. The network established an end-to-end mapping of the time-dependent skin surface temperature after short-pulse laser irradiation to the temperature gradient along the skin depth, allowing vascular parameters to be extracted directly from the observed data measured by the infrared camera. The model trained by data containing noise of multiple intensities could identify and attenuate noise in the original data, which is beneficial to suppress the broadening effect of a significant broadening of reconstructed temperature peaks induced by noise and significantly improve the reconstruction accuracy of less than 5%. In comparison with traditional iterative methods, the new model shows technical superiority in its unparalleled accuracy and low computational cost as well as great potential for real-time monitoring of subcutaneous vascular features.