An integrated system from microscopy to AI for real-time object detection in endometrial cytology

Endometrial cytology, which is minimally invasive and available as an outpatient procedure, is widely used in Japan for early detection of endometrial cancer, but its diagnostic process is time-consuming and requires expert diagnosticians. We developed a real-time artificial intelligence (AI)-assisted system using a standard microscope, a charge-coupled device (CCD) camera, and the You-Only-Look-Once version 5x (YOLOv5x) (a well-established object detection model) to support endometrial cytology screening in resource-limited settings. A total of 146 pre-operative cytology cases were collected, and the model was trained to detect abnormal cell clusters. The system was evaluated in real-time using a CCD camera, and its diagnostic performance was compared with that of three pathologists and four medical students. In an independent test of 20 cases, the AI model achieved an accuracy of 85%, showing promising performance comparable to the average accuracy of 75% among human evaluators. Furthermore, the median diagnostic time was reduced by approximately 45% with AI assistance. The impact of AI support varied by user expertise, with notable improvements among non-specialists. This proof-of-concept study demonstrates the feasibility and potential of affordable, real-time AI support for endometrial cytology using widely available equipment. Further validation with larger, multicenter datasets is warranted to confirm the generalizability and clinical utility of this approach.