Table 1.

Comparison of proposed end-to-end learning-based normalization and unmixing compared to the benchmark dual-band normalization followed by non-negative least squares unmixing

	Dual-band	ACU-Net	ACU-SA	PLS	MLP
Phantom data	R = 0.93 RMSE = 3.77	R = 0.997 RMSE = 0.19	R = 0.98 RMSE = 0.51	R = 0.93 RMSE = 0.35	R = 0.998 RMSE = 1.31
Pig brain homogenate	R = 0.82 RMSE = 4.17	R = 0.99 RMSE = 0.33	R = 0.91 RMSE = 0.81	R = 0.67 RMSE = 2.10	R = 0.92 RMSE = 1.94

PLS and MLP approaches are also compared. The coefficient of determination between known and computed PpIX concentration is used for consistency with previous normalization work.^61,67 For human data, no labels are available, so the reconstruction’s MSE (ReMSE) is used.