Fig. 3.

Schematic of the proposed deep convolutional neural network, i.e., the FDU-Net. FC: fully connected layers. $L_{roi}$: prior-weighted loss function; ${\text{y}}^{\text{truth}}$: ground truth 3D optical image volume of ${\mu }_a$; $y^{inter}$: intermediate model outputs upon completing the first training step; ${\text{y}}^{\text{final}}$: eventual optical image recovered after the second training step.