Abstract
Endometrial cytology is essential in detecting and diagnosing endometrial cancer, a prevalent gynecological malignancy. However, diagnosis proves intricate and time-intensive due to hormone-induced morphological changes in cells and cell cluster thickness. While recent artificial intelligence (AI)-supported cytodiagnosis systems rely on whole-slide imaging (WSI), focusing issues caused by cell cluster thickness hinder endometrial cytological slide digitalization. Despite the high demand for these systems, progress in endometrial cytodiagnosis has been slow. This study utilizes You Only Look Once (YOLOv5x) under a microscope to detect abnormal cell clusters in real-time without the need for WSI. We analyzed 146 preoperative endometrial cytology cases collected at Nippon Medical School between 2017 and 2023, confirmed by hysterectomy specimens. And we trained YOLOv5x using 3,151 images captured with a smartphone from 96 cytology slides. For real-time detection, images were captured via a microscope-mounted charge-coupled device (CCD) camera and processed by YOLOv5x. For real-time abnormal evaluation, thresholds for cell cluster and slide levels were adjusted using 30 new cases. The AI model's diagnoses for 20 new cases were compared with those made by pathologists and medical students with varying experience levels. The AI model outperformed human evaluators, achieving accuracy, precision, and recall of 85%, 82%, and 90%, respectively. Additionally, AI-assisted diagnosis shortened the median evaluation time from 4,458 to 2,460 seconds, equivalent to a reduction of 44.82%. Although diagnosis accuracy of inexperienced medical students did not significantly improve, notable enhancements in recall were achieved among pathologists and AI-trained students, particularly those familiar with the AI system. Overall, our findings demonstrate that the proposed AI system significantly hastens detection of abnormal cell clusters while seamlessly integrating into existing workflows without the need for expensive specialized equipment, which makes it particularly suitable for resource-constrained settings.