Atrial fibrillation (AF) is the most common supraventricular arrhythmia and its most specific feature on the electrocardiogram (ECG) is the presence of fibrillatory waves (F-waves). The aim of this study is to present a new method that innovatively uses deep-learning (DL) as a filter to optimize the extraction of F-waves from ECGs. To do so, the CPSC database, containing 918 12-lead ECGs showing normal sinus rhythm (NSR), and Reference database, containing 30 12-lead ECG created by combining real F-waves and QRST complexes, were used. Zero vectors and the real F-waves were used as ground truth to evaluate the method. ECGs were segmented into 1-second windows, that represent the inputs of the method. The DL method comprises two convolutional neural networks having the same architecture (six sequential multipath modules), but different loss functions. The root mean squared error (RMSE) between amplitudes of the estimated and ground truth F-waves was computed, together with the area under the curve (AUC) of the receiver operating characteristics. Results indicate a low testing RMSE (NSR: 6.02 µV; AF: 11.14 µV) and a high testing AUC (>99%). In conclusion, our DL method can reliably extract F-waves from ECGs; their estimated amplitude permits reliable discrimination of AF patients.