Wang Jinli, Tong Jin, Li Jun, Cao Chunli, Wang Sirui, Bi Tianyu, Zhu Peishan, Shi Linan, Deng Yaqian, Ma Ting, Hou Jixue, Cui Xinwu
Department of Ultrasound, the First Affiliated Hospital of Medical College, Shihezi University, Shihezi, China.
School of Business Administration, Lanzhou University of Finance and Economics, Lanzhou, China.
Quant Imaging Med Surg. 2024 Oct 1;14(10):7111-7127. doi: 10.21037/qims-24-679. Epub 2024 Sep 26.
Breast cancer is one of the most common malignancies in women worldwide, and early and accurate diagnosis is crucial for improving treatment outcomes. Conventional ultrasound (CUS) is a widely used screening method for breast cancer; however, the subjective nature of interpreting the results can lead to diagnostic errors. The current study sought to estimate the effectiveness of using a GoogLeNet deep-learning convolutional neural network (CNN) model to identify benign and malignant breast masses based on CUS.
A literature search was conducted of the Embase, PubMed, Web of Science, Wanfang, China National Knowledge Infrastructure (CNKI), and other databases to retrieve studies related to GoogLeNet deep-learning CUS-based models published before July 15, 2023. The diagnostic performance of the GoogLeNet models was evaluated using several metrics, including pooled sensitivity (PSEN), pooled specificity (PSPE), the positive likelihood ratio (PLR), the negative likelihood ratio (NLR), the diagnostic odds ratio (DOR), and the area under the curve (AUC). The quality of the included studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies Scale (QUADAS). The eligibility of the included literature were independently searched and assessed by two authors.
All of the 12 studies that used pathological findings as the gold standard were included in the meta-analysis. The overall average estimation of sensitivity and specificity was 0.85 [95% confidence interval (CI): 0.80-0.89] and 0.86 (95% CI: 0.78-0.92), respectively. The PLR and NLR were 6.2 (95% CI: 3.9-9.9) and 0.17 (95% CI: 0.12-0.23), respectively. The DOR was 37.06 (95% CI: 20.78-66.10). The AUC was 0.92 (95% CI: 0.89-0.94). No obvious publication bias was detected.
The GoogLeNet deep-learning model, which uses a CNN, achieved good diagnostic results in distinguishing between benign and malignant breast masses in CUS-based images.
乳腺癌是全球女性中最常见的恶性肿瘤之一，早期准确诊断对于改善治疗效果至关重要。传统超声（CUS）是一种广泛应用的乳腺癌筛查方法；然而，解读结果的主观性可能导致诊断错误。本研究旨在评估使用谷歌神经网络深度学习卷积神经网络（CNN）模型基于CUS识别乳腺良恶性肿块的有效性。
对Embase、PubMed、Web of Science、万方、中国知网（CNKI）等数据库进行文献检索，以获取2023年7月15日前发表的与基于谷歌神经网络深度学习CUS模型相关的研究。使用包括合并敏感度（PSEN）、合并特异度（PSPE）、阳性似然比（PLR）、阴性似然比（NLR）、诊断比值比（DOR）和曲线下面积（AUC）等多种指标评估谷歌神经网络模型的诊断性能。使用诊断准确性研究质量评估量表（QUADAS）评估纳入研究的质量。纳入文献的合格性由两位作者独立检索和评估。
所有12项以病理结果作为金标准的研究均纳入荟萃分析。敏感度和特异度的总体平均估计值分别为0.85 [95%置信区间（CI）：0.80 - 0.89]和0.86（95% CI：0.78 - 0.92）。PLR和NLR分别为6.2（95% CI：3.9 - 9.9）和0.17（95% CI：0.12 - 0.23）。DOR为37.06（95% CI：20.78 - 66.10）。AUC为0.92（95% CI：0.89 - 0.94）。未检测到明显的发表偏倚。
使用CNN的谷歌神经网络深度学习模型在基于CUS的图像中区分乳腺良恶性肿块方面取得了良好的诊断结果。
