M dwarfs are the most common stars in the Milky Way, and their motion brings to light the evolution of the Milky Way. Radial velocity (RV) of M dwarf is one of the most important parameters to reflect its motion. Chinese large scale scientific construction project LAMOST survey has obtained hundreds of thousands of M-type stars and their radial velocities measurement which needs the automatic and efficient program. The most commonly used method of measuring the radial velocity of an M star is to cross correlation between its observed spectra and the template spectra. However, in the actual process, idiosyncrasies of individual spectra and noises often lead to the loss of accuracy of radial velocities. To alleviate the effect induced by the above factors, in this paper, we try to improve radial velocities measurement of M dwarfs especially for those spectra with higher signal noise ratio but with local noises. A statistical method combining experienced features is proposed to select discriminative wavelength ranges of spectra for radial velocity measurement, avoiding noise affected wavelength regions. Some spectra from LAMOST DR3 M-type catalogue are used to test this method and the results are compared with the radial velocities of the same target from APOGEE. The experimental results show that the proposed method can effectively improve the accuracy of radial velocity measurement for M-type stars by reducing the influence of local spectral noises.