腹腔镜下人工智能自动视觉识别子宫内膜异位症病变的初步结果:一项概念验证研究
这项多中心概念验证研究收集了2020年1月至2023年8月间112例疑似子宫内膜异位症患者的腹腔镜手术视频,训练基于YOLOv5的深度神经网络识别九类病变。模型对「表浅黑色」「表浅细微」和「卵巢巧克力样液体」三类表现良好(F1分数分别为0.94、0.74、0.75),对「致密粘连」「卵巢子宫内膜瘤」和「深部子宫内膜异位症」可接受(F1分数分别为0.70、0.63、0.632),对「表浅红色」「表浅白色」和「膜状粘连」表现较弱。研究结论认为该技术可行,但需进一步改进以提高准确性和精确度。
要点
- 基于YOLOv5的深度神经网络可识别腹腔镜下九类子宫内膜异位症病变。
- 「表浅黑色」类F1分数最高(0.94),「膜状粘连」类最低(0.16和0.02,原文如此)。
- 模型在每段视频至少一帧中识别多数病变的潜力较强,但需进一步改进。
核心结论
- 本研究证明人工智能用于腹腔镜手术中子宫内膜异位症视觉识别是可行的。
- 初步结果令人鼓舞,但需进一步开发以提高模型性能并标准化标注方法。
- 人工智能融入外科实践有望辅助子宫内膜异位症诊断并改善手术结局。
研究设计与样本
- 多中心概念验证研究,收集并标注手术视频,训练、验证和测试深度神经网络。
- 手术视频来自法国、匈牙利、巴西和丹麦的专家中心。
- 收集2020年1月至2023年8月间112例因疑似子宫内膜异位症接受腹腔镜手术患者的视频序列。
- 纳入可识别子宫内膜异位症病变的序列,排除图像质量差和既往手术操作过的序列。
主要发现
- 模型对「表浅黑色」「表浅细微」和「卵巢巧克力样液体」类的F1分数分别为0.94、0.74和0.75,表现良好。
- 模型对「致密粘连」「卵巢子宫内膜瘤」和「深部子宫内膜异位症」类的F1分数分别为0.70、0.63和0.632,表现可接受。
- 模型对「表浅红色」「表浅白色」和「膜状粘连」类的F1分数分别为0.25、0.18、0.16和0.02,表现较弱(原文如此列出四个数值)。
临床怎么用
- 作者认为人工智能融入外科实践有望辅助子宫内膜异位症诊断并改善手术结局。
- 作者指出需进一步开发以提高模型性能并标准化标注方法。
局限性
- 原文指出需进一步改进以提高准确性和精确度。
- 原文指出需进一步开发以提高模型性能并标准化标注方法。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To develop a machine learning method for the automatic recognition of endometriosis lesions during laparoscopic surgery and evaluate its feasibility and performance. DESIGN: Collecting and annotating surgical videos and training, validating, and testing a deep neural network. SETTING: Multicenter proof-of-concept study using surgical videos from expert centers in France, Hungary, Brazil, and Denmark. PARTICIPANTS: Surgical video sequences were collected from 112 patients who underwent laparoscopic procedures for suspected endometriosis between January 2020 and August 2023. Sequences with identifiable endometriosis lesions were included, while poor-quality images and sequences with prior surgical manipulation were excluded. INTERVENTIONS: A deep neural network based on YOLOv5 was trained to detect and classify nine visual classes of endometriosis lesions (superficial black, superficial red, superficial white, superficial subtle, filmy adhesions, dense adhesions, deep endometriosis, ovarian endometrioma, and ovarian chocolate fluid). RESULTS: The model performance was good for the "superficial black," "superficial subtle," and "ovarian chocolate fluid" classes (F1 score = 0.94, 0.74, and 0.75, respectively), acceptable for the "dense adhesion," "ovarian endometrioma" and "deep endometriosis" classes (F1 score = 0.70, 0.63 and 0.632, respectively), and weak for the "superficial red," "superficial white," and "filmy adhesions" classes (F1 score = 0.25, 0.18, 0.16 and 0.02, respectively). However, while these results highlight the model's strong potential in identifying most lesions in at least one frame of each sequence, they underscore the need for further refinement to improve accuracy and precision. CONCLUSION: This study demonstrates the feasibility of applying artificial intelligence for visual recognition of endometriosis during laparoscopic surgery. While the initial results are encouraging, further development is needed to enhance the model performance and standardize the annotation methods. The integration of AI in surgical practice holds promise for assisting in endometriosis diagnosis and improving surgical outcomes.
引用
- Netter A, Noorzadeh S, Duchateau F, et al. Initial Results in the Automatic Visual Recognition of Endometriosis Lesions by Artificial Intelligence During Laparoscopy: A Proof-of-Concept Study. J Minim Invasive Gynecol. 2025. PMID: 40912397. doi:10.1016/j.jmig.2025.08.027
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。