Fluorescence imaging during to oral cancer surgery is typically 2D, yielding limited information on tumor depth. Here, we continue the development of a spatial frequency domain imaging (SFDI) system for 3D fluorescence imaging. A deep convolutional neural network takes as inputs SFDI-computed absorption, scattering and spatial-frequency fluorescence images, and yields images of fluorescence concentration and tumour depth. The model is trained using in silico data from Monte Carlo simulations of geometric tumor shapes (e.g., cylinder, spherical harmonics). Initial results yield average depth errors of <0.1 mm. Experiments are conducted in agar phantoms based on patient imaging.
|