Accurate channel estimation plays a major role in the performance of orthogonal frequency division multiplexing (OFDM). The channel estimation performance depends on how effectively the channel state information can be extracted using known pilot symbols. The placement of these pilot symbols influences both channel estimation and subsequent detection performance, and the conventional fixed comb- or block-type pilot arrangements may not always provide the best results in time-varying channel conditions. Therefore, this work presents a Convolutional Neural Network (CNN)-based pilot position selection approach. In the proposed work, we generate a set of various time-varying channel realizations under six different 3GPP channel conditions and utilize the generated dataset to train the CNN. We evaluate these channel realizations using different pilot patterns, enabling the trained CNN to select pilot positions effectively in time-varying channel conditions. To evaluate the effectiveness of the proposed learning-based pilot position selection, we compare the results with Particle Swarm Optimization (PSO)-based pilot position selection and fixed diamond pilot arrangement methods. The simulation results indicate that both the CNN- and PSO-based pilot arrangements achieve better Bit Error Rate (BER) and Normalized Mean Squared Error (NMSE) than the diamond pilot arrangement, with the CNN achieving slightly better performance than the PSO-optimized one. Therefore, a trained CNN is a promising approach for adaptive pilot position selection without using an iterative approach under time-varying channel conditions.