Table 1.

Quantitative performance comparison of the different Beer–Lambert models and network training strategies on the test set of the two spectral datasets.

Dataset	Individual ID	Spectral MAE		Concentration MAE
		Non-scattering	Scattering	Strategy (a)	Strategy (b)
Broadband NIRS	507	$1.23\times {10}^{-2}$	$\mathbf{8.60}\times {\mathbf{10}}^{-\mathbf{3}}$	$1.36\times {10}^{-2}$	$\mathbf{4.32}\times {\mathbf{10}}^{-\mathbf{3}}$
	509	$1.09\times {10}^{-2}$	$\mathbf{9.37}\times {\mathbf{10}}^{-\mathbf{3}}$	$2.46\times {10}^{-2}$	$\mathbf{4.81}\times {\mathbf{10}}^{-\mathbf{3}}$
	511	$8.01\times {10}^{-3}$	$\mathbf{6.60}\times {\mathbf{10}}^{-\mathbf{3}}$	$1.16\times {10}^{-2}$	$\mathbf{3.41}\times {\mathbf{10}}^{-\mathbf{3}}$
	512	$1.20\times {10}^{-2}$	$\mathbf{1.09}\times {\mathbf{10}}^{-\mathbf{2}}$	$1.25\times {10}^{-2}$	$\mathbf{4.59}\times {\mathbf{10}}^{-\mathbf{3}}$
HELICoiD	012-01	$3.27\times {10}^{-2}$	$\mathbf{2.49}\times {\mathbf{10}}^{-\mathbf{2}}$	$1.73\times {10}^{-1}$	$\mathbf{1.64}\times {\mathbf{10}}^{-\mathbf{2}}$
	012-02	$2.24\times {10}^{-2}$	$\mathbf{2.19}\times {\mathbf{10}}^{-\mathbf{2}}$	$1.50\times {10}^{-1}$	$\mathbf{2.58}\times {\mathbf{10}}^{-\mathbf{2}}$
	015-01	$6.33\times {10}^{-2}$	$\mathbf{2.54}\times {\mathbf{10}}^{-\mathbf{2}}$	$1.75\times {10}^{-1}$	$\mathbf{1.54}\times {\mathbf{10}}^{-\mathbf{2}}$

To compare the two (non-scattering and scattering) models, we compute the mean absolute error of the spectral fit (denoted as “Spectral MAE”) between the GT observed and predicted signals. The two network training strategies are compared by assessing the MAE of each strategy between the network and optimization-inferred concentrations (denoted as “Concentration MAE”) of all considered chromophores. In the case of the HELICoiD dataset, only pixels labeled as normal, tumor, or blood were considered for these computations. The best-performing model and strategy for each individual is highlighted in bold.