In order to achieve a fast, accurate, and robust reconstruction of fluorescence molecular tomography (FMT), limited-projection FMT and permissible region selection methods have been drawing more and more attention. Aiming to solve the problems of existing permissible region selection methods including the difficulty of parameter setting and the inaccuracy of multi-objective selection, so as to improve reconstruction quality of the limited-projection FMT, a method through which permissible regions were selected by applying an iterative self-organizing data analysis technique algorithm (ISODATA) was proposed. Firstly, ISODATA was used to cluster the primary reconstruction results, and then permissible regions were selected at every separated cluster. A contrast experiment of reconstructing three target fluorophores was designed to verify the feasibility and effectiveness in the application. The experimental results indicate that positions of all the three fluorophores can be reconstructed accurately with two projections only by using the proposed selection method. With four projections, the average localization error of reconstruction results is 0.18 mm and the relative error of the fluorescence yield is less than 50 %. Meanwhile, the threshold method fails to reconstruct and the relative error of the fluorescence yield using the region-shrinking method is 61.2 %. The proposed method is able to select permissible regions accurately and efficiently even with little measurement data. Consequently, the accuracy and robustness of limited-projection FMT reconstruction are improved.