Aprendizagem de máquina para correção do espectro de emissão de nanopartículas usadas para nanotermometria em aplicações biológicas

Abstract

In the field of nanothermometry, the use of nanoparticle (NP) emissions has proven to be an important tool in various areas, particularly for the early diagnosis of cancer and localized inflammations. This is because regions with these characteristics exhibit temperatures different from normal tissue. Although the use of NP emissions as nanothermometers is a very promising area, there are several obstacles to overcome, such as spectral distortions caused by tissues. This work used Machine Learning (ML) for spectral correction in tissue data using the absorption coefficient µa (brain gray matter, premenopausal breast, liver, skin, skin with varying water concentration, and water) and the scattering coefficient (µs) of brain gray matter, premenopausal breast, liver, and skin. The algorithms that achieved the best results in spectral correction were Decision Tree, Random Forest, and k-Nearest Neighbors. On average, they obtained MAE, MSE, MAPE, R², and σ values among the best possible. The Decision Tree stood out the most due to its low time cost and for supporting the hypothesis, using the Wilcoxon test, that its predictive capacity for this scenario achieved accurate and precise results. Each of the spectral corrections was also used to determine the temperature of the tissues. The models that stood out the most were Decision Tree, Random Forest, and k-Nearest Neighbors. The temperature violin plot of each of these predictors showed accuracy and precision, correctly determining the temperature to one decimal place.