Abstract：
OBJECTIVE: To develop a multi-instance learning (MIL) based artificial intelligence (AI)-assisted diagnosis models by using laryngoscopic images to differentiate benign and malignant vocal fold leukoplakia (VFL).
METHODS: The AI system was developed, trained and validated on 5362 images of 551 patients from three hospitals. Automated regions of interest (ROI) segmentation algorithm was utilized to construct image-level features. MIL was used to fusion image level results to patient level features, then the extracted features were modeled by seven machine learning algorithms. Finally, we evaluated the image level and patient level results. Additionally, 50 videos of VFL were prospectively gathered to assess the system\'s real-time diagnostic capabilities. A human-machine comparison database was also constructed to compare the diagnostic performance of otolaryngologists with and without AI assistance.
RESULTS: In internal and external validation sets, the maximum area under the curve (AUC) for image level segmentation models was 0.775 (95 % CI 0.740-0.811) and 0.720 (95 % CI 0.684-0.756), respectively. Utilizing a MIL-based fusion strategy, the AUC at the patient level increased to 0.869 (95 % CI 0.798-0.940) and 0.851 (95 % CI 0.756-0.945). For real-time video diagnosis, the maximum AUC at the patient level reached 0.850 (95 % CI, 0.743-0.957). With AI assistance, the AUC improved from 0.720 (95 % CI 0.682-0.755) to 0.808 (95 % CI 0.775-0.839) for senior otolaryngologists and from 0.647 (95 % CI 0.608-0.686) to 0.807 (95 % CI 0.773-0.837) for junior otolaryngologists.
CONCLUSIONS: The MIL based AI-assisted diagnosis system can significantly improve the diagnostic performance of otolaryngologists for VFL and help to make proper clinical decisions.
METHODS: The AI system was developed, trained and validated on 5362 images of 551 patients from three hospitals. Automated regions of interest (ROI) segmentation algorithm was utilized to construct image-level features. MIL was used to fusion image level results to patient level features, then the extracted features were modeled by seven machine learning algorithms. Finally, we evaluated the image level and patient level results. Additionally, 50 videos of VFL were prospectively gathered to assess the system\'s real-time diagnostic capabilities. A human-machine comparison database was also constructed to compare the diagnostic performance of otolaryngologists with and without AI assistance.
RESULTS: In internal and external validation sets, the maximum area under the curve (AUC) for image level segmentation models was 0.775 (95 % CI 0.740-0.811) and 0.720 (95 % CI 0.684-0.756), respectively. Utilizing a MIL-based fusion strategy, the AUC at the patient level increased to 0.869 (95 % CI 0.798-0.940) and 0.851 (95 % CI 0.756-0.945). For real-time video diagnosis, the maximum AUC at the patient level reached 0.850 (95 % CI, 0.743-0.957). With AI assistance, the AUC improved from 0.720 (95 % CI 0.682-0.755) to 0.808 (95 % CI 0.775-0.839) for senior otolaryngologists and from 0.647 (95 % CI 0.608-0.686) to 0.807 (95 % CI 0.773-0.837) for junior otolaryngologists.
CONCLUSIONS: The MIL based AI-assisted diagnosis system can significantly improve the diagnostic performance of otolaryngologists for VFL and help to make proper clinical decisions.
摘要：
目的：通过使用喉镜图像来区分良性和恶性声带白斑（VFL），开发基于多实例学习（MIL）的人工智能（AI）辅助诊断模型。
方法：开发了人工智能系统，对来自三家医院的551名患者的5362张图像进行了培训和验证。利用自动感兴趣区域（ROI）分割算法来构建图像级特征。MIL用于将图像级别结果融合到患者级别特征，然后利用七种机器学习算法对提取的特征进行建模。最后,我们评估了图像水平和患者水平结果.此外,前瞻性收集了50个VFL视频，以评估系统的实时诊断能力。还构建了人机比较数据库，以比较有和没有AI辅助的耳鼻喉科医师的诊断性能。
结果：在内部和外部验证集中，图像水平分割模型的最大曲线下面积（AUC）为0.775（95％CI0.740-0.811）和0.720（95％CI0.684-0.756），分别。利用基于MIL的融合策略，患者水平的AUC增加至0.869(95%CI0.798-0.940)和0.851(95%CI0.756-0.945).对于实时视频诊断，患者水平的最大AUC达到0.850(95%CI,0.743-0.957).在AI的帮助下,高级耳鼻喉科医师的AUC从0.720(95%CI0.682-0.755)提高到0.808(95%CI0.775-0.839),初级耳鼻喉科医师的AUC从0.647(95%CI0.608-0.686)提高到0.807(95%CI0.773-0.837).
结论：基于MIL的AI辅助诊断系统可以显着提高耳鼻喉科医师对VFL的诊断能力，并有助于做出正确的临床决策。
方法：开发了人工智能系统，对来自三家医院的551名患者的5362张图像进行了培训和验证。利用自动感兴趣区域（ROI）分割算法来构建图像级特征。MIL用于将图像级别结果融合到患者级别特征，然后利用七种机器学习算法对提取的特征进行建模。最后,我们评估了图像水平和患者水平结果.此外,前瞻性收集了50个VFL视频，以评估系统的实时诊断能力。还构建了人机比较数据库，以比较有和没有AI辅助的耳鼻喉科医师的诊断性能。
结果：在内部和外部验证集中，图像水平分割模型的最大曲线下面积（AUC）为0.775（95％CI0.740-0.811）和0.720（95％CI0.684-0.756），分别。利用基于MIL的融合策略，患者水平的AUC增加至0.869(95%CI0.798-0.940)和0.851(95%CI0.756-0.945).对于实时视频诊断，患者水平的最大AUC达到0.850(95%CI,0.743-0.957).在AI的帮助下,高级耳鼻喉科医师的AUC从0.720(95%CI0.682-0.755)提高到0.808(95%CI0.775-0.839),初级耳鼻喉科医师的AUC从0.647(95%CI0.608-0.686)提高到0.807(95%CI0.773-0.837).
结论：基于MIL的AI辅助诊断系统可以显着提高耳鼻喉科医师对VFL的诊断能力，并有助于做出正确的临床决策。
