This paper introduces an orthogonal moment pre-processing method to enhance convolutional neural network outcomes for whole brain image segmentation in magnetic resonance images. The method implements kernel windows based on orthogonal moments to transform the original image into a modified version with orthogonal moment properties. The transformed image contains the optimal representation of the coefficients of the Legendre, Tchebichef and Pseudo-Zernike moments. The approach was evaluated on three distinct datasets; NFBS, OASIS, and TCIA, and obtained an improvement of 4.12%, 1.91%, and 1.05%, respectively. A further investigation employing transfer learning using orthogonal moments of various orders and repetitions, achieved an improvement of 9.86% and 7.76% on the NFBS and OASIS datasets, respectively, when trained using the TCIA dataset. In addition, the best image representations were used to compare different convolutional neural network architectures, including U-Net, U-Net++, and U-Net3+. U-Net3+ demonstrated a slight improvement over U-Net in an overall accuracy of 0.64 % for the original image and 0.33 % for the modified orthogonal moment image. Statement of SignificanceThis manuscript introduces a method to initialize convolutional neural network using orthogonal moment filters for whole brain image segmentation in magnetic resonance images. Three orthogonal moments were selected and tests were performed in three distinct datasets. Also, the comparison of three different convolutional neural network (U-Net, U-Net++, and U-Net3+) were conducted. The use of an initial orthogonal moment filter for convolutional neural network in brain segmentation in magnetic resonance imaging achieved an improvement over conventional method. The findings in this study contribute to the long-standing search for the development of a pre-processing technique for whole brain segmentation in MRI.