Hybrid attention-enhanced MobileNetV2 with particle swarm optimization for endometrial cancer classification in CT images

Endometrial cancer is a form of uterine cancer that is known to be deadly and shows a strong therapeutic response if diagnosed at an early stage. The inability of traditional endometrial cancer methods to provide timely and cost-effective diagnosis has been transformed with the introduction of computational techniques driven by oncologists and data scientists. Deep learning, the most important branch of artificial intelligence, has found increasing importance in diagnosing endometrial cancer. This paper presents a novel methodology for accurate diagnosis of endometrial cancer computed tomography (CT) images, based on the use of a hybrid deep learning framework to develop a novel methodology that automates hyperparameter optimization and enhances feature recognition by integrating dual attention and particle swarm optimization (PSO) techniques. The pre-trained MobileNetV2 backbone uses geometric transformations (rotations, translations, and reflections) while extracting hierarchical features from CT slices to mitigate data scarcity. PSO is used to enhance the hyperparameters governing the attention and regularization modules. The method combines efficient swarm-based optimization and adaptive attention mechanisms, improving the discrimination between different images and establishing a reproducible pipeline for medical imaging applications with less illustrative data. The performance of the model was validated using a new dataset, collected from King Abdullah University Hospital in Jordan by physicians, and the proposed model achieved an accuracy of 86.07%, a precision of 86.75%, a sensitivity of 86.02%, a specificity of 91.45%, and an AUC of 97.33%. , outperforming all previously trained models (MobileNetV2, VGG16, VGG19, ResNets50, NASNetMobile, and InceptionV3), on the King Abdullah University Hospital Endometrial Cancer Computed Tomography (KAUH-ECCTD) dataset. PSO optimization enabled effective tuning of key hyperparameters (learning rate, dropout rate, L2 regularization, number of neurons), directly enhancing model generalization and discrimination capability. The validated model, trained on a dataset collected from King Abdullah University Hospital (KAUH-ECCTD), has strong potential for real-world clinical applications as part of AI-assisted diagnostic tools and clinical decision support systems for oncologists. The proposed approach can enhance early detection, personalized treatment planning, and continuous monitoring in endometrial cancer management, thereby facilitating collaborative research between oncologists, biomedical engineers, and data scientists.