Attention-enhanced RT-DETR-T for cervical cell detection

Cervical cancer screening remains highly reliant on manual cytology, which is subjective and error-prone. Traditional two-stage methods combining segmentation and classification often suffer from error propagation. To address this, we propose RT-DETR-T, a detection network that directly identifies abnormal cervical cells without explicit segmentation. The model incorporates a Self-Regulating Triplet Attention and FPN-based feature fusion into the HGNetV2 backbone, enhancing fine-grained feature representation. Furthermore, a Center-Aware Inclusive Localization Loss enforces region coverage and center alignment for robust detection. Experiments show that RT-DETR-T achieves high recall and accurate localization, offering a promising solution for automated cervical cancer screening.