基于深度学习的卵巢肿块良恶性分类及泛化性能研究OA

Objective To evaluate the diagnostic performance of deep learning models on both representative and randomly selected clinical ultrasound images,and to compare their performance with that of radiologists with different levels of experience,in order to objectively assess the potential of artificial intelligence-assisted diagnosis in real-world clinical settings.Methods A total of 936 ultrasound images from 168 patients,retrospectively collected from December 2022 to July 2024 at the Quanzhou First Hospital Affiliated to Fujian Medical University(Quanzhou First Hospital,Fujian),were included in this study.The dataset was divided into training,validation,and test sets in a ratio of 7:1:2.Six convolutional neural network models(VGG19_bn,DenseNet121,Swin Transformer-Tiny,ConvNeXt-Tiny,MobileNetV2,and ResNet101)were employed to evaluate diagnostic performance on representative images selected by radiologists.To further assess model generalization,a fully random image sampling strategy was applied,and the experiments were repeated.The diagnostic performance of the best-performing model was quantitatively compared with that of three radiologists with senior,intermediate,and junior levels of experience.Results On representative images,the VGG19_bn model achieved the best performance(AUC=0.906).In the random image testing scenario,DenseNet121 demonstrated the strongest robustness(AUC=0.888),outperforming the senior ultrasound radiologist(AUC=0.738).Notably,DenseNet121 showed a superior balance between sensitivity(0.786)and specificity(0.857).In contrast,radiologist diagnosis tended to exhibit high sensitivity but relatively low specificity,with variability in sensitivity observed across different experience levels.Conclusion Well-trained deep learning models not only achieve strong diagnostic performance under controlled conditions but also demonstrate stable generalization ability in randomly sampled,clinically realistic scenarios.Their overall diagnostic performance is comparable to that of radiologists,highlighting their potential value in real-world clinical applications.