We performed three-dimensional magnetic particle imaging using third-harmonic signal detection. The measurement system consists of three coil types, i.e., an ac excitation coil, a dc gradient coil, and five pickup coils cooled to 77 K. In the experiment, two magnetic nanoparticle samples containing 10 μg (Fe) of magnetic nanoparticles were arranged three-dimensionally with a spacing of 16 mm. An ac excitation field with a strength of 1.6 mT at 3 kHz and a dc gradient field with a field gradient of 0.2 T/m were used. A third-harmonic signal generated from the magnetic nanoparticle samples was detected using the five pickup coils. The noise of the detection system was as low as SB1/2 = 7 fT/Hz1/2 at a signal frequency of 9 kHz. A magnetic field map generated from the two magnetic nanoparticle samples was obtained by scanning the samples two-dimensionally. Field maps obtained with the five pickup coils were analyzed to reconstruct a three-dimensional distribution image of the magnetic nanoparticle sample. A mathematical technique called the nonnegative least squares method was used for this purpose. We successfully demonstrated three-dimensional magnetic particle imaging. Namely, we reconstructed the three-dimensional positions of the two magnetic nanoparticle samples with good accuracy. Quantities of the two magnetic nanoparticle samples were also estimated reasonably well. These results indicate the feasibility of the present system for three-dimensional magnetic particle imaging.