Seismic imaging requires a dense configuration of receivers and sources to obtain high-quality interpretable data for hydrocarbon exploration. Nonetheless, economic and environmental conditions commonly limit the number of sources. Current seismic data acquisition methods focus on computationally reconstructing or interpolating missing sources, reducing the costs. However, most works are based on sparse optimization, where the quality depends on the sparsity level, or on deep learning, where the limited availability of training data also limit the training process. Therefore, this work proposes an internal learning framework for reconstructing missing sources, addressing the lack of training data issue. Simulations show that the proposed approach outperforms state-of-the-art methods with gains of up to 12 dB in the PSNR and 11 dB in SNR metrics.