Department of Radiology, Stanford University, Stanford, California, USA; Deusto Institute of Technology, University of Deusto/Ikerbasque, Basque Foundation for Science, Bilbao, Spain.
Department of Radiology, Stanford University, Stanford, California, USA.
Ultrasound Med Biol. 2022 Oct;48(10):2060-2078. doi: 10.1016/j.ultrasmedbio.2022.05.031. Epub 2022 Jul 30.
Adiposity accumulation in the liver is an early-stage indicator of non-alcoholic fatty liver disease. Analysis of ultrasound (US) backscatter echoes from liver parenchyma with deep learning (DL) may offer an affordable alternative for hepatic steatosis staging. The aim of this work was to compare DL classification scores for liver steatosis using different data representations constructed from raw US data. Steatosis in N = 31 patients with confirmed or suspected non-alcoholic fatty liver disease was stratified based on fat-fraction cutoff values using magnetic resonance imaging as a reference standard. US radiofrequency (RF) frames (raw data) and clinical B-mode images were acquired. Intermediate image formation stages were modeled from RF data. Power spectrum representations and phase representations were also calculated. Co-registered patches were used to independently train 1-, 2- and 3-D convolutional neural networks (CNNs), and classifications scores were compared with cross-validation. There were 67,800 patches available for 2-D/3-D classification and 1,830,600 patches for 1-D classification. The results were also compared with radiologist B-mode annotations and quantitative ultrasound (QUS) metrics. Patch classification scores (area under the receiver operating characteristic curve [AUROC]) revealed significant reductions along successive stages of the image formation process (p < 0.001). Patient AUROCs were 0.994 for RF data and 0.938 for clinical B-mode images. For all image formation stages, 2-D CNNs revealed higher patch and patient AUROCs than 1-D CNNs. CNNs trained with power spectrum representations converged faster than those trained with RF data. Phase information, which is usually discarded in the image formation process, provided a patient AUROC of 0.988. DL models trained with RF and power spectrum data (AUROC = 0.998) provided higher scores than conventional QUS metrics and multiparametric combinations thereof (AUROC = 0.986). Radiologist annotations indicated lower hepatic steatosis classification accuracies (Acc = 0.914) with respect to magnetic resonance imaging proton density fat fraction that DL models (Acc = 0.989). Access to raw ultrasound data combined with artificial intelligence techniques may offer superior opportunities for quantitative tissue diagnostics than conventional sonographic images.
肝脏内的脂肪堆积是非酒精性脂肪肝的早期指标。使用深度学习 (DL) 分析肝实质的超声 (US) 背向散射回波可能为肝脂肪变性分期提供一种负担得起的替代方法。本工作旨在比较使用不同数据表示形式从原始 US 数据构建的用于肝脂肪变性的 DL 分类评分。使用磁共振成像作为参考标准，根据脂肪分数截断值对 N = 31 例确诊或疑似非酒精性脂肪性肝病患者进行脂肪变性分层。采集 US 射频 (RF) 帧（原始数据）和临床 B 型图像。从 RF 数据中模拟中间图像形成阶段。还计算了功率谱表示和相位表示。使用配准的补丁独立训练 1-D、2-D 和 3-D 卷积神经网络 (CNN)，并比较分类评分与交叉验证。2-D/3-D 分类有 67,800 个补丁可用，1-D 分类有 1,830,600 个补丁可用。结果还与放射科医师 B 型模式注释和定量超声 (QUS) 指标进行了比较。补丁分类评分（接收者操作特征曲线下的面积 [AUROC]）显示在图像形成过程的连续阶段显著降低（p < 0.001）。RF 数据的患者 AUROC 为 0.994，临床 B 型图像的患者 AUROC 为 0.938。对于所有图像形成阶段，2-D CNN 的补丁和患者 AUROC 均高于 1-D CNN。使用功率谱表示训练的 CNN 比使用 RF 数据训练的 CNN 收敛速度更快。在图像形成过程中通常丢弃的相位信息提供了患者 AUROC 为 0.988。使用 RF 和功率谱数据训练的 DL 模型（AUROC = 0.998）提供的评分高于传统的 QUS 指标及其组合（AUROC = 0.986）。放射科医师注释表明，与基于磁共振成像质子密度脂肪分数的肝脂肪变性分类准确性（Acc = 0.914）相比，DL 模型（Acc = 0.989）的准确性较低。与传统超声图像相比，使用人工智能技术访问原始超声数据可能为定量组织诊断提供更好的机会。
