Deep Learning Estimation of Absorbed Dose for Nuclear Medicine Diagnostics

In radionuclide therapy with ^{177}Lu, the absorbed-dose distribution can be approximated by convolving the time-integrated activity distribution with a dose voxel kernel for a single tissue type. This approximation is fast but inaccurate: it treats the body as homogeneous and therefore ignores the tissue heterogeneity that governs where energy is deposited. The heterogeneity can be recovered by combining computed tomography and single-photon emission computed tomography with a Monte Carlo transport simulation, at a high computational cost. We investigate whether the map from a local density kernel to the corresponding dose voxel kernel can instead be learned from data by a convolutional neural network, so that density-adapted kernels become available without a full transport calculation for each patient. On held-out patient data, the proposed U-residual architecture reaches a continuous intersection-over-union score of 0.86 after 308 epochs, with a mean squared error of 1.24\times 10^{-4} on the normalised targets. This generalisation to unseen data indicates that the network approximates, rather than merely memorises, the simulation-based map from density kernels to dose voxel kernels. The network does not replace the underlying transport physics; it approximates the density-to-dose association that a full Monte Carlo transport simulation would otherwise have to supply anew for every patient.