本研究基于肝脏超声射频（RF）信号与超声衰减参数（UAP），以肝脏穿刺活检病理结果为金标准，构建ConvNeXt深度学习模型，探讨其对肝病显著肝脂肪变性的诊断效能。

回顾性分析2020年1月至2025年2月于首都医科大学附属北京地坛医院就诊的1222例肝病患者的RF、UAP和病理数据，将数据集划分为训练集851例、验证集192例与测试集179例。于肝脏穿刺活检前行RF信号获取和UAP测量，并收集患者肝穿刺病理结果。将病理肝脂肪变性F0~F1级归为阴性（标签为0），≥F2级归为阳性（标签为1）。为构建显著肝脂肪变性评估的人工智能模型，对每一列超声RF信号数据进行Z-score标准化处理，以ConvNeXt-Tiny作为主干网络，采用其在ImageNet数据集上的预训练参数进行初始化，并将首层三通道权重的均值作为单通道权重。以UAP作为特征标量，将其与ConvNeXt主干网络提取的图像特征在最终线性层进行拼接，经过全连接层与Sigmoid函数输出模型预测得分，通过阈值判定完成最终分类。绘制ConvNeXt深度学习模型及UAP判定肝病显著肝脂肪变性（≥F2级）的ROC曲线，并比较二者的诊断效能。

本研究ConvNeXt深度学习模型预测得分诊断肝病显著肝脂肪变性的ROC曲线下面积在测试集上为0.838（95%置信区间：0.769~0.906），在验证集确定的诊断阈值（0.225）下，测试集的混淆矩阵[[真阴性（TN）假阳性（FP）][假阴性（FN）真阳性（TP）]] 为[[10836][629]]，敏感度为82.86%，特异度为75.00%。此外，根据测试集ROC曲线计算的最佳截断值为0.272时，混淆矩阵[[TN FP] [FN TP]] 为[[11529][629]]，敏感度为82.86%，特异度为79.86%。UAP诊断显著肝脂肪变性的ROC曲线下面积在测试集上为0.802（95%置信区间：0.723~0.881），取诊断阈值269 dB/m时，敏感度为74.29%，特异度为72.22%。

基于RF信号与UAP构建的ConvNeXt深度学习模型在识别肝病显著肝脂肪变性（≥F2级）中具有一定的性能优势，有望发展成为临床筛查和随访脂肪性肝病患者的有效工具。

To develop a ConvNeXt deep learning model based on hepatic ultrasound radiofrequency (RF) signals and the ultrasound attenuation parameter (UAP), using liver biopsy as the reference standard, and to evaluate its diagnostic performance for significant hepatic steatosis in liver diseases.

A total of 1222 patients with liver diseases who underwent evaluation at Beijing Ditan Hospital, Capital Medical University between January 2020 and February 2025 were retrospectively analyzed. RF signals, UAP measurements, and liver biopsy data were collected. The dataset was divided into a training set (n=851), a validation set (n=192), and a test set (n=179). RF signal acquisition and UAP measurement were performed prior to liver biopsy, and corresponding histopathological findings were obtained. Hepatic steatosis graded as F0–F1 was defined as negative (label 0), while ≥F2 was defined as positive (label 1). To construct an artificial intelligence model for assessing significant hepatic steatosis, each column of ultrasound RF signal data was normalized using Z-score standardization. ConvNeXt-Tiny was used as the backbone network, initialized with pretrained weights from the ImageNet dataset, and the mean of the first-layer three-channel weights was used to initialize single-channel weights. UAP was incorporated as a scalar feature and concatenated with image features extracted by the ConvNeXt backbone at the final linear layer. The model prediction score was generated through a fully connected layer followed by a Sigmoid function, and final classification was determined using a threshold. Receiver operating characteristic (ROC) curves were generated to evaluate the performance of the ConvNeXt model and UAP for detecting significant hepatic steatosis, and their performance was compared.

In the test set, the area under the ROC curve (AUC) of the ConvNeXt model for diagnosing significant hepatic steatosis was 0.838 (95% confidence interval [CI]: 0.769–0.906). Using the cutoff value (0.225) determined in the validation set, the confusion matrix [[True Negative (TN) False Positive (FP) ][False Negative (FN)True Positive (TP)]] in the test set was [[108 36][6 29]], yielding a sensitivity of 82.86% and specificity of 75.00%. In addition, based on the ROC curve in the test set, the optimal cutoff value was 0.272, at which the confusion matrix [[TN FP ][FN TP]] was [[115 29][6 29]], yielding a sensitivity of 82.86% and specificity of 79.86%. For UAP, the AUC for diagnosing significant hepatic steatosis was 0.802 (95%CI: 0.723–0.881), with a threshold of 269 dB/m, and the sensitivity and specificity were 74.29% and 72.22%, respectively.

The ConvNeXt deep learning model based on RF signals and UAP demonstrates favorable performance in identifying significant hepatic steatosis (≥F2) in liver diseases, and may serve as a useful tool for clinical screening and follow-up of patients with steatotic liver disease.