Predicting the thermophysical properties of skin tumor based on the surface temperature and deep learning

Predicting the thermophysical properties of the skin tumor is a great challenge in the field of biomedical engineering, which is helpful for the diagnostic of the tumor. In this paper, the relationship between thermophysical properties of the tumor and the time-dependent skin surface temperature could be revealed through dynamic thermography and deep learning. The deep learning model for the inverse bio-heat conduction problem is used to identify the overall thermophysical properties of the skin tumor, including the depth, size, thermal conductivity, heat generation and blood perfusion of the skin tumor. Firstly, a 3D numerical skin model with different layers, including the tumor, muscle, fat, dermis and epidermis, is constructed to calculate the surface temperature under different thermophysical properties of the skin tumor. And the numerical model is verified by comparing the time-dependent skin surface temperature of Clark II and Clark IV tumors. Then the deep learning model is established to relate the time-dependent surface temperature with the thermophysical properties and trained by the numerical simulation data. The performances of the deep learning model are examined by the Clark II and Clark IV tumors with different measurement errors. The results show that the deep learning model can learn the abstract features of the time-dependent surface temperature and estimate the tumor properties by the skin surface temperature. Compared with the Clark IV tumor, the measurement errors have more influence on the Clark II tumor. At last, the influences of seven thermophysical properties of the tumor on the skin surface temperature are further numerically analyzed to understand the deep learning model predictions. Interestingly, it is found that the deep learning model can well predict the tumor heat generation and blood perfusion of the skin tumor. The numerical simulation results show that the surface temperature profiles are influenced by the properties mentioned. However, the normalized temperature variation profiles do not. The proposed method provides a useful diagnostic tool for estimating the thermophysical properties of the skin tumor. (c) 2021 Elsevier Ltd. All rights reserved.