Fig. 3From: Concept development of an on-chip PET systemTraining pipeline of the scintillation position prediction. The light patterns recorded on each surface are stacked along the channels dimension and used as input to the CNN. The CNN should predict the gamma-ray interaction position from the stacked light patterns. The MAE loss is computed with the predicted and ground truth positions and backpropagated through the network to update its weightsBack to article page