Precision localization of the anus using a smart defecation assistance system based on the YOLOv11n model

BackgroundExisting defecation-assistance devices generally lack automated anal localization and image-based screening for conditions that may preclude probe operation.ObjectiveTo develop and evaluate an image-based method for anal localization and abnormal-region detection in an intelligent defecation-assistance system.MethodsThe dataset comprised 1600 human images from 1000 patients and 500 animal images. YOLOv11n and YOLOv11s were trained using transfer learning. A mixed-species technical test set was used for model comparison and ablation analyses, whereas final performance was evaluated on 160 human images. Precision, recall, F1-score, mAP, inference speed, latency, and GPU memory usage were assessed.ResultsOn the mixed-species technical test set, YOLOv11n achieved an F1-score of 97.1%, mAP@0.5 of 99.1%, and 67.1 FPS. On the human-only final evaluation set, it achieved an F1-score of 96.5% and mAP@0.5 of 98.6%. Removing data augmentation and animal images reduced mAP@0.5 by 2.1 and 1.8 percentage points, respectively. Edge-device latency was 14.9 ms per frame, with 2.8 GB GPU memory usage.ConclusionYOLOv11n showed high performance and real-time inference on static images, supporting its technical feasibility for probe localization and abnormality screening. Clinical safety and effectiveness require prospective patient evaluation.