Traditional numerical methods of computing radiative transfer problems, both forward and inverse, are computationally expensive, and machine learning methods hold the promise of significantly reducing computational costs while maintaining accuracy. Especially for inverse problems, machine learning methods can be used to improve the accuracy of the results, and, in some cases, provide solutions where traditional methods fail. In this study, a machine learning method with neural networks is used to solve the inverse problem of diffuse optical tomography (DOT), a medical imaging method that constructs non-invasive optical images for biological tissue. Since light propagation in biological tissue is highly scattering, traditional DOT image reconstruction methods tend to produce low-quality images with the problem of over-smoothing. We proposed a neural network model, which contains several neural networks to solve the inverse problem in DOT. The case considered in the study is to detect circular objects in a scattering-absorbing medium. The results sought are location, size, and optical properties (LSOP). The result shows the proposed LSOP model is more accurate and less time-consuming compared to traditional image reconstruction methods, with significant improvement in the accuracy of reconstructed results for most cases. The image correlation coefficient (ICC) between the original and the reconstructed images improves from the 0.5-0.9 range for traditional methods to 0.9-0.99 for machine learning, with lower values for very small objects or objects with optical properties very close to those of the background.