Corrección de atenuación en SPECT mediante aprendizaje profundo con ventanas de fotopico y dispersión

Attenuation artifacts represent a major limitation in the interpretation of SPECT studies. Although CT-based attenuation correction remains the reference method, it requires dedicated infrastructure and involves additional radiation exposure. In this work, we propose a deep learning model based on a dual-input U-Net capable of generating synthetic attenuation correction maps from the photopeak and scatter windows, while incorporating body contour segmentation. A total of 320 SPECT myocardial perfusion studies using 99mTc-MIBI were included: 258 for training and validation, and 62 for evaluation. The proposed model achieved an MSSIM of 0.98 and an MRE of 7.28% for the attenuation maps, and an MSSIM of 0.99 and an MRE of 4.8% for the reconstructed images. The 95% limits of agreement were [−6.87, 6.76] % at the voxel level and [−7.81, 8.21] % at the segment level. These results indicate that combining photopeak information, the scatter window, and body segmentation provides an emission-only alternative with the potential to avoid the use of CT for attenuation correction.