This study presents a novel application approach to automated segmentation of endometrial cancer using deep learning techniques in PET and CT images using a single model architecture to help improve the segmentation results and assist medical practitioners in the diagnosis of endometrial cancer. Motivation: Endometrial cancer, the leading gynecologic cancer in developed countries, requires accurate segmentation for diagnosis and proper treatment planning. We propose an endometrial cancer segmentation method based on a modified DeepLabV3+ model applied to multimodality PET/CT imaging data Methodology: The proposed methodology utilizes a single model architecture for PET and CT modalities, simplifying computational complexity with the guarantee of high segmentation results. The pipeline utilizes a pre-trained ResNet50 encoder from ImageNet, an Atrous Spatial Pyramid Pooling layer, and a decoder tailored to estimate accurate boundaries. BCE with logit loss and the Adam optimizer were utilized for 20-epoch training of the model. Dataset: We utilized the ECPC-IDS dataset, the largest and only publicly available open-source multimodal endometrial cancer segmentation dataset. Results: The results are superior to previous approaches, which reported Dice values of 0.8161 and 0.7780 for PET and CT modalities, respectively, and consistent IoU scores or Jaccard coefficients greater than 0.64. Conclusion: This paper advances computer-aided diagnosis systems for endometrial cancer and shows the efficiency of deep learning techniques for medical image segmentation problems. Future work includes 1) model expansion for 3D volumetric segmentation 2) application of an end-to-end multimodal fusion to combine both PET and CT information, 3) using attention mechanisms to further improve the feature representation and boundary accuracy, and 4) model validation on larger multicenter datasets to evaluate generalizability.