EJNMMI Physics

Table 7 Parameters used for every experiment run

From: Concept development of an on-chip PET system

Parameter	Value
Dataset Splits	80k/800k train, 20k/200k val, 10k test
Input Tensor Size	Cx32x32/Cx24x24/Cx16x16
Network Architecture	ResNet18
Loss Function	Mean Absolute Error (MAE)
Optimizer	Adam [29]
Learning Rate	3e−4
Batch Size	256
Steps	50,000/500,000
Float Precision	16 bit

The number of input channels C is determined by the number of surfaces with SiPMs

Back to article page