基于深度学习的后天原发性中耳胆脂瘤及中耳乳突炎CT图像分类模型的应用研究

Application of Deep Learning-Based CT Image Classification Models for Acquired Cholesteatoma and Otitis Media

目的 研究基于深度学习(deep learning, DL)的中耳胆脂瘤及中耳乳突炎人工智能分类诊断模型,评估其教学及临床应用价值,探讨其提高诊断效率和准确性的潜力。方法 回顾性分析2021年1月—2023年8月于中国人民解放军总医院第六医学中心治疗的200例中耳疾病患者,包括100例后天原发性中耳胆脂瘤患者和100例中耳乳突炎患者。所有患者术前均接受了颞骨高分辨率CT(high-resolution computed tomography,HRCT)检查,并通过外科手术及病理学证实诊断。从上述患者的HRCT图像中选取具有病灶特征性层面的1000张CT图像建立数据集,按照6:1:3的比例随机分为训练组(n=600),验证组(n=100)和测试组(n=300)。使用3种先进的医学图像分类模型—Convolutional Neural Networks Meet Vision Transformers(CMT)模型、Efficient Vision Transformer模型、CrossShaped Window模型,进行模型训练与效能评估,最终在测试集进行医学图像分类测试。选择准确率最高者作为本研究的最优模型。最后与初级组、中级组、高级组不同年资的临床医师组进行图像分类结果比较,评价人工智能模型的诊断效能。采用χ^(2)检验进行统计学分析,检验标准P=0.0125。结果 CMT模型作为本研究的最优模型,其诊断的准确率、精确度、敏感度、特异度分别为90.0%、90.0%、90.7%、89.3%。CMT模型诊断准确率优于初级医师组但低于高级临床医师,与中级医师接近。初级组、中级组、高级组临床医师组阅片时间低于人工智能模型。结论 深度学习诊断模型具有一定鉴别中耳胆脂瘤及中耳乳突炎的能力,具有良好的诊断效能。

Objective To develop and validate a deep learning(DL)based artificial intelligence(AI)diagnostic model for middle ear cholesteatoma and mastoiditis,as well as its potential applications to enhance diagnostic efficiency and accuracy.Methods A retrospective analysis was conducted on data from 200 patients treated for middle ear diseases at the Department of Otolaryngology,Sixth Medical Center of the PLA General Hospital,from January 2021 to August 2023,including 100 with acquired primary middle ear cholesteatoma and 100 with mastoiditis.All patients underwent preoperative high-resolution computed tomography(HRCT)of the temporal bone,and diagnoses were confirmed by surgical and/or pathological findings.1000 HRCT images featuring characteristic lesion changes were randomly selected for training(n=600),validation(n=100),or testing(n=300).The Convolutional Neural Networks Meet Vision Transformers(CMT),Efficient Vision Transformer,and Cross-Shaped Window models were employed for model training,efficacy evaluation,and testing of best-performing models,respectively.The AI models performance was compared with those by junior,intermediate and senior level clinicians.Chi-square tests were used for statistical analysis,with a significance threshold of P=0.0125.Results The CMT model emerged as the optimal model,achieving an accuracy(ACC)of 90.0%,precision(PRE)of 90.0%,sensitivity(SEN)of 90.7%,and specificity(SPE)of 89.3%.The diagnostic accuracy of the CMT model surpassed that by junior level clinicians but was lower than that by senior level clinicians,and comparable to that by mid-level clinicians.In terms of image review time,all clinician groups were faster than the AI model.Conclusion The deep learning based CMT model demonstrates a significant capability to differentiate middle ear cholesteatoma from mastoiditis,offering robust diagnostic performance.