Numerical simulations were used to test the performance of the proposed algorithm. A three-dimensional (3-D) breast simulation was created from MRI slices of an actual patient with a tumor size of . The corresponding DCE (b) and DW (c) images used to construct the and matrices are shown in Fig. 2. The MRI slices were manually segmented into tumor, adipose, and fibroglandular regions based on the preconstrast T1 and DCE images. After segmentation, typical HbT, water, SA, and SP values for tumor (, 70.5%, 1.23, and 0.26), adipose (, 47%, 1.17, and 0.2), and fibrogandular (, 60%, 1.2, and 0.23) tissues from Ref. 12 were assigned to each region, as shown in Fig. 2(d). For simulation, 16 colocated source-detector positions were placed around the breast. For each source illumination, data were collected at 15 detector locations. Thus, the total number of measurements was 2160 () involving nine wavelengths (from 661 to 948 nm). Amplitude-dependent Gaussian noise with a variance of 1% was added to the simulated measurements. To obtain the initial estimates of the optical properties, the simulated data set was calibrated with a known homogeneous circular phantom. Images were reconstructed using a pixel basis of . The same and of 0.01 and 0.4 were used to construct and , respectively. The initial parameters used in different methods were same, and their value was 100.