Accurate segmentation of myocardial edema and scar is essential for diagnosing and managing cardiac diseases. While each imaging sequence or modality offers unique advantages, logistical constraints make it challenging to obtain annotated data for every modality. This study introduces a novel method called GenSegNet, which combines an end-to-end 3D UNet with a generative adversarial network (GAN) and self-training to enhance segmentation accuracy. The GAN framework generates missing sequences, while pseudo labels from unlabeled data are used to fine-tune the segmentation model, further improving its performance. To address slice correlation, we trained an end-to-end 3D UNet using three MRI sequences—late gadolinium enhancement (LGE), T2-weighted (T2), and balanced steady-state free precession (bSSFP)—to segment scar, edema, left ventricle, and myocardium. Our model showed promising results on the validation set from the \(MyoPS^{++}\) challenge, securing 4th place among the participating teams. Compared to nnU-Net, GenSegNet demonstrated a notable improvement, particularly with a 3.48% increase in edema segmentation precision and an 8.97% increase in scar segmentation precision.