轻量化YOLO模型在非标准体表肿瘤图像检测中的性能研究

Performance Study of Lightweight YOLO Model for Non-Standard Superficial Tumor Image Detection

摘要: 体表肿瘤作为整形外科学和皮肤病学的重要研究对象，其早期检测对改善患者预后至关重要。本研究提出了一种基于非标准化体表肿瘤图像的智能诊断框架，结合YOLO系列目标检测模型(v7至v10)，针对十类常见体表肿瘤实现高效筛查。数据集来源于青岛大学附属医院整形外科的非标准临床图像，涵盖多样化的光照、设备和背景条件，模拟真实场景。实验结果表明，YOLOv10n在检测性能上表现最佳(F1分数0.912，mAP@0.5 0.912，总推理时间4.3 ms)，YOLOv8n以0.952的卓越精度超越Faster R-CNN和EfficientDet等传统模型。尽管数据分布不均和图像变异性对稀有类别(如蓝痣)检测构成挑战，YOLO框架仍展现出较强的鲁棒性与实时性。本研究为非标准化场景下的体表肿瘤自动检测提供了技术支持，其轻量化设计适配智能手机等低成本设备，有望推动远程筛查应用，改善体表肿瘤早期诊断效率与患者预后。

Abstract: The early detection of superficial tumors is an important research object in plastic surgery and dermatology, which is crucial for improving patient prognosis. In this study, we propose an intelligent diagnostic framework based on non-standardized superficial tumor images by YOLO series of target detection models (v7 to v10) for ten types of common superficial tumors. The dataset was derived from non-standard clinical images from the Department of Plastic and Reconstructive Surgery of the Affiliated Hospital of Qingdao University, covering diverse lighting, equipment, and background conditions to simulate real-life scenarios. Experimental results indicated that YOLOv10n performs the best detection (F1-score 0.912, mAP@0.5 0.912, total inference time 4.3 ms). Additionally, YOLOv8n surpasses conventional models, including Faster R-CNN and EfficientDet, with exceptional accuracy (0.952). Despite the uneven distribution of the data and the image variability, which present challenges for rare category (blue mole) detection, the hybrid YOLO framework demonstrates robustness and real-time performance. This study provides technical support for automated superficial tumor detection in non-standardized scenarios. Its lightweight design is suitable for low-cost devices, including smartphones, which can promote remote screening applications and improve the efficiency of early diagnosis of superficial tumors and patient prognosis.