Positron probes can accurately localize malignant tumors by directly detecting positrons emitted from positron-emitting radiopharmaceuticals that accumulate in malignant tumors. In the conventional method for direct positron detection, multilayer scintillator detection and pulse shape discrimination techniques are used. However, some γ-rays cannot be distinguished by conventional methods. Accordingly, these γ-rays are misidentified as positrons, which may increase the error rate of positron detection. To analyze the energy distribution in each scintillator of the multilayer scintillator detector to distinguish true positrons and γ-rays and to improve the positron detection algorithm by discriminating true and false positrons. We used Autoencoder, an unsupervised deep learning architecture, to obtain the energy distribution data in each scintillator of the multilayer scintillator detector. The Autoencoder was trained to separate the combined signals generated from the multilayer scintillator detector into two signals of each scintillator. An energy window was then applied to the energy distribution obtained using the trained Autoencoder to distinguish true positrons from false positrons. Finally, the performance of the proposed method and conventional positron detection algorithm was evaluated in terms of the sensitivity and error rate for positron detection. The energy distribution map obtained using the trained Autoencoder was proven to be similar to that of the simulated results. Furthermore, the proposed method demonstrated a 29.79% (+0.42%p) increase in positron detection sensitivity compared to the conventional method, both having an equal error rate of 0.48%. However, when both methods were set to have the same sensitivity of 1.83%, the proposed method had an error rate that was 25.0% (-0.16%p) lower than that of the conventional method. We proposed and developed an Autoencoder-based positron detection algorithm that can discriminate between true and false positrons with a smaller error rate than conventional methods. We verified that the proposed method could increase the positron detection sensitivity while maintaining a low error rate compared to the conventional method. If the proposed algorithm is implemented in handheld positron detection probes or cameras, diseases such as cancers can be more accurately localized in a shorter time compared with using traditional methods.