中国中西医结合影像学杂志

目的：本研究旨在评估弱监督深度学习模型与强监督深度学习模型在评估高级别和低级别透明细胞肾细胞癌（ccRCC）中的表现。方法：共纳入经病理确诊的ccRCC患者300例，数据按7：3的比例分为训练集（210例）和测试集（90例）。强监督和弱监督深度学习模型分别基于3D ResNet-18网络和2D ResNet-18网络对CT增强扫描的3期图像（皮髓质期、肾实质期和平扫）进行深度学习，最终形成6个深度学习模型。计算每个模型的准确度、敏感度、特异度和AUC用以评估深度学习模型的判别效果。结果：基于测试集3期影像的弱监督深度学习模型中，单张切片模型与≥3层切片模型诊断性能相当，且皮髓质期影像的单张切片模型诊断性能最高，其准确率、敏感度、特异度和AUC值分别为0.859、0.857、0.860和0.907，与强监督深度学习皮髓质期影像模型（准确率、敏感度、特异度和AUC值分别为0.848、0.854、0.843和0.922）相比，差异无统计学意义（P>0.05）。结论：本研究利用皮髓质期图像开发的弱监督深度学习模型在区分高级别和低级别ccRCC方面具有与强监督深度学习模型相同的高诊断性能。

Objective: This study aimed to evaluate the performance of weakly-supervised deep learning (DL) models versus strongly-supervised DL models in distinguishing high-grade from low-grade clear cell renal cell carcinoma (ccRCC). Methods: A total of 300 patients with pathologically confirmed ccRCC were enrolled， and the data were divided into a training set (210 cases) and a test set (90 cases) in a 7:3 ratio. Strongly-supervised and weakly-supervised DL models were developed using a 3D ResNet-18 network and a 2D ResNet-18 network， respectively， analyzing triphasic contrast-enhanced CT images (CMP， NP， and UP phases). This resulted in six distinct DL models. The discriminative performance of each model was assessed by calculating accuracy， sensitivity， specificity， and AUC. Results: In the weakly supervised deep learning model of three-phase images in the test set， the diagnostic performance of the single slice model is comparable to that of the ≥ 3-layer slice model， and the single slice model of the corticomedullary phase image has the highest diagnostic performance， with accuracy， sensitivity， specificity， and AUC values of 0.859， 0.857， 0.860， and 0.907， respectively. Compared with the best-performing strongly supervised deep learning model using corticomedullary phase images (with accuracy， sensitivity， specificity， and AUC values of 0.848， 0.854， 0.843， and 0.922， respectively)， the differences were not statistically significant (P > 0.05). Conclusion: The weakly-supervised DL model developed using CMP phase images demonstrated high diagnostic performance equivalent to that of strongly-supervised DL models in differentiating high-grade from low-grade ccRCC.