Luminescence-enabled three-dimensional temperature bioimaging

Abstract

Luminescence thermometry affords remote thermal readouts with high spatial resolution in a minimally invasive way. This technology has advanced our understanding of biological mechanisms and physical processes from the macro- to the submicrometric scale. Yet, current approaches only allow obtaining 2D thermal images. This aspect limits the potential of this technology, given the inherent three-dimensional nature of heat diffusion processes. Despite initial attempts, a credible method that allows extracting 3D thermal images via luminescence is missing. Here, we design such a method combining Ag₂S nanothermometers and machine learning algorithms. The approach leverages the distortions in the emission spectra of luminescent nanothermometers caused by changes in temperature and tissue-induced photon extinction. The optimized neural network-based algorithm can extract this information and provide 3D thermal images of complex nanothermometer patterns. Although tested for luminescence thermometry at the in vivo level, this method has far-reaching implications for luminescence-supported 3D sensing in biological systems in general.

Fig. 1. Ag₂S nanothermometers (NTs) optical properties.

a Ag₂S NTs were selected for this study because of their absorption in NIR-I, and the overlap of the emission spectrum with the absorption peak of water falling in NIR-II. The expected variations in the spectrum of Ag₂S NTs due to temperature and tissue thickness (depth) changes are schematically reported in (b) and (c), respectively.

Fig. 2. Calibration of the nanothermometers (NTs) and training of the algorithm for three-dimensional (3D) luminescence thermometry.

a Scheme of the experimental setup used to calibrate the luminescent NTs as a function of temperature (T), and tissue thickness (d). b Hyperspectral image (integrated intensity over the 1000–1400 nm range) obtained with T = 29 °C and d = 1 mm (phantom with 0.1% Intralipid content) alongside the spectrum extracted from a single pixel of the image. The scale bar is 2 mm. c Exemplary normalized spectra calibration dataset for different values of T and d. d Scheme of the convolutional neural networks (CNN) + dense neural networks (DNN) algorithm used in this study. e Predicted-vs-real values of T and d obtained from the trained algorithm using phantom and real tissues (i.e., chicken breast, pork muscle, cow liver). These results were obtained by randomly analyzing 20% of the calibration dataset. The temperature (T) range of 26 to 45 °C includes data numbers 4108, 4219, 4240, 4138, 4225, 4110, 4210, 4224, 4226, respectively. The thickness (d) range of 1 to 6 mm includes data numbers 6191, 6350, 6373, 6274, 6343, 6165, respectively. In those violin representations, the central white point is the mean value, the black box encompasses the 25–75% range of the data, the black lines represent the 1.5 IQR (interquartile range), and the kernel density estimates are modeled using the Kernel Smooth function in Origin®.

Fig. 3. Test of three-dimensional (3D) thermal imaging capabilities.

a Scheme of the phantom with inserted capillaries with Ag₂S nanothermometers (capillary w. Ag₂S NTs, outer diameter (OD) = 1 mm, inner diameter (ID) = 0.5 mm). b Top view of the phantom and (c) corresponding hyperspectral image (integrated intensity over the 1000–1400 nm range) used to obtain the 3D thermal images shown in (d) after analysis by convolutional neural network (CNN) and dense neural network (DNN). The uncertainties, calculated from the standard deviations, are based on data points from the lower (larger z-value) to the higher capillary, with sample sizes of 196, 165, and 193, respectively. e Thermal images obtained considering the depth measured from a photo (distance between the top surface of the phantom and the capillary axis) and the temperature resulting from simulations. f Scheme of the real tissue composite with inserted Ag₂S NTs-filled capillaries (OD = 1, ID = 0.5 mm). g Top view of the tissue composite and (h) corresponding hyperspectral image (integrated intensity over the 1000–1400 nm range) used to obtain the 3D thermal images shown in (i). The number of data points for the lower and higher capillaries were 664 and 513, respectively. j Thermal images obtained considering the depth measured from a photo (distance between the top surface of the tissue composite and the capillary axis) and the temperature resulting from simulations. All scale bars are 5 mm.

Fig. 4. Three-dimensional (3D) thermal imaging in vivo.

a Representation of an anaesthetized nude mouse to which Ag₂S nanothermometers (NTs, 100 μL, 8 mg mL⁻¹) are injected retro-orbitally to obtain a thermal readout of ventral vasculature. b Hyperspectral image (integrated intensity over the 1000–1400 nm range) of the ventral vasculature of the mouse obtained under 808-nm excitation, alongside the binarized image used as a mask to select the pixel to be analyzed are reported. c 3D thermal image of the vasculature obtained from the analysis of the hyperspectral image in (b). The projections of the experimental points along the xy and xz planes are also reported in gray.