Chen Jianan, Zhang Weibin, Bao Jingwen, Wang Kun, Zhao Qiannan, Zhu Yuli, Chen Yanling
The First Affiliated Hospital of Guangdong Pharmaceutical University, Guangzhou, Guangdong, China.
Department of Ultrasound, Huashan Hospital, Fudan University, Shanghai, China.
Abdom Radiol (NY). 2024 Jan;49(1):93-102. doi: 10.1007/s00261-023-04089-4. Epub 2023 Nov 24.
The current study developed an ultrasound-based deep learning model to make preoperative differentiation among hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and combined hepatocellular-cholangiocarcinoma (cHCC-ICC).
The B-mode ultrasound images of 465 patients with primary liver cancer were enrolled in model construction, comprising 264 HCCs, 105 ICCs, and 96 cHCC-ICCs, of which 50 cases were randomly selected to form an independent test cohort, and the rest of study population was assigned to a training and validation cohorts at the ratio of 4:1. Four deep learning models (Resnet18, MobileNet, DenseNet121, and Inception V3) were constructed, and the fivefold cross-validation was adopted to train and validate the performance of these models. The following indexes were calculated to determine the differential diagnosis performance of the models, including sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), F-1 score, and area under the receiver operating characteristic curve (AUC) based on images in the independent test cohort.
Based on the fivefold cross-validation, the Resnet18 outperformed other models in terms of accuracy and robustness, with the overall training and validation accuracy as 99.73% (± 0.07%) and 99.35% (± 0.53%), respectively. Furthers validation based on the independent test cohort suggested that Resnet 18 yielded the best diagnostic performance in identifying HCC, ICC, and cHCC-ICC, with the sensitivity, specificity, accuracy, PPV, NPV, F1-score, and AUC of 84.59%, 92.65%, 86.00%, 85.82%, 92.99%, 92.37%, 85.07%, and 0.9237 (95% CI 0.8633, 0.9840).
Ultrasound-based deep learning algorithm appeared a promising diagnostic method for identifying cHCC-ICC, HCC, and ICC, which might play a role in clinical decision making and evaluation of prognosis.
本研究开发了一种基于超声的深度学习模型，用于在术前鉴别肝细胞癌（HCC）、肝内胆管癌（ICC）和肝细胞-胆管细胞癌（cHCC-ICC）。
将465例原发性肝癌患者的B超图像纳入模型构建，其中包括264例HCC、105例ICC和96例cHCC-ICC，随机选取50例组成独立测试队列，其余研究人群按4:1的比例分为训练队列和验证队列。构建了4种深度学习模型（Resnet18、MobileNet、DenseNet121和Inception V3），并采用五折交叉验证来训练和验证这些模型的性能。计算以下指标以确定模型的鉴别诊断性能，包括基于独立测试队列图像的敏感性、特异性、准确性、阳性预测值（PPV）、阴性预测值（NPV）、F-1分数和受试者操作特征曲线下面积（AUC）。
基于五折交叉验证，Resnet18在准确性和稳健性方面优于其他模型，总体训练和验证准确率分别为99.73%（±0.07%）和99.35%（±0.53%）。基于独立测试队列的进一步验证表明，Resnet 18在识别HCC、ICC和cHCC-ICC方面具有最佳诊断性能，敏感性、特异性、准确性、PPV、NPV、F1分数和AUC分别为84.59%、92.65%、86.00%、85.82%、92.99%、92.37%、85.07%和0.9237（95%CI 0.8633，0.9840）。
基于超声的深度学习算法似乎是一种用于识别cHCC-ICC、HCC和ICC的有前景的诊断方法，可能在临床决策和预后评估中发挥作用。
