探讨多模态超声联合人工智能S-Detect技术校正乳腺影像报告与数据系统（BI-RADS）分类对诊断乳腺肿块良恶性的价值。

本研究首先采用常规超声、超微血流成像技术及应变弹性成像技术，将2021年7月至12月安徽医科大学第二附属医院收集的连续130例乳腺肿块病例作为训练集进行超声检查，超微血流成像及弹性成像结果分别以血管指数（VI）、弹性应变率（SR）值表示，以病理结果为金标准得出良恶性肿块VI值、SR值的截断值；然后以2022年1月至5月连续110例乳腺肿块作为验证集联合人工智能S-Detect技术，采用常规超声进行BI-RADS分级诊断，再以超微血管成像技术、应变弹性成像技术及S-Detect技术评估结果校正BI-RADS分级，以病理结果为金标准绘制受试者操作特征（ROC）曲线，采用Z检验比较不同诊断方法（常规超声+S-Detect+VI值+ SR值联合诊断以及各方法独立诊断）ROC曲线下面积的差异，计算不同诊断方法的敏感度、特异度、准确性、阳性预测值和阴性预测值。

训练集130例乳腺肿块中恶性70例、良性60例，VI值及SR值良恶性截断值分别为4.05、2.59。验证集110例乳腺肿块中恶性63例、良性47例，常规超声、S-Detect、VI值、SR值及四者联合诊断乳腺肿块良恶性的ROC曲线下面积分别为0.936、0.588、0.827、0.802、0.785，联合诊断的效能优于单独应用各独立模块，差异具有统计学意义（Z=6.074，P＜0.001；Z=2.668，P=0.008；Z=3.084，P=0.002；Z=3.293，P=0.001），联合诊断的敏感度为98.4%、特异度为87.2%、准确性为93.6%、阳性预测值为91.2%、阴性预测值为97.6%。根据2013版美国放射学会BI-RADS≥4类肿块应行穿刺活检，穿刺活检率由87.3%（96/110）降至61.8%（68/110），并校正4例被错判为良性的恶性病例（非特殊类型的浸润性乳腺癌3例，导管内原位癌1例），校正32例错判为恶性的良性病例（腺病17例、腺病伴纤维腺瘤14例、叶状肿瘤1例）。

多模态超声联合人工智能S-Detect技术校正BI-RADS分类可提升乳腺肿块良恶性的诊断效能，减少不必要的穿刺活检、提高乳腺恶性肿块的检出率。

To explore the value of multimodal ultrasonography in correcting breast imaging reporting and data system (BI-RADS) classification of benign and malignant breast masses.

Conventional ultrasound, ultramicro blood flow imaging, and strain elastography were used to examine 130 consecutive cases of breast masses collected from the Second Affiliated Hospital of Anhui Medical University from July 2021 to December 2021 as a training set. The results of ultramicro blood flow imaging and elastography are expressed as vascular index (VI) and elastic strain (SR) ratio, respectively. Taking the pathological results as the gold standard, the cut-off values of VI and SR for the diagnosis of benign and malignant masses were obtained. Then, 110 consecutive cases of breast masses collected from January to May 2022 were selected as the verification set, BI-RADS grading was performed by conventional ultrasound, and the BI-RADS grading results were corrected by the evaluation results of ultramicroangiography, strain elastography, and artificial intelligence based S-Detect technology. The receiver operating characteristic (ROC) curve was drawn according to the pathological results. The area under the ROC curve (AUC) values of combined diagnosis and independent diagnosis were compared by Z test, and the sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of different diagnostic methods were calculated.

Among the 130 cases of breast masses in the training set, 70 were malignant and 60 were benign. The cut-off values of VI and SR were 4.05 and 2.59, respectively. Among the 110 cases of breast masses in the verification set, 63 were malignant and 47 were benign; the AUC values of conventional ultrasound, S-Detect, VI, SR, and their combination in the diagnosis of benign and malignant breast masses were 0.936, 0.588, 0.827, 0.802, and 0.785, respectively, and Z test showed that the efficacy of combined diagnosis was better than that of any independent modality (Z=6.074, P＜0.001; Z=2.668, P=0.008; Z=3.084, P=0.002; Z=3.293, P=0.001). For joint diagnosis, the sensitivity was 98.4%, the specificity was 87.2%, the accuracy was 93.6%, the PPV was 91.2%, and the NPV was 97.6%. According to the 2013 version of American College of Radiology BI-RADS, for ≥ category 4 masses, puncture biopsy should be performed. After correction, the puncture biopsy rate decreased from 87.3% (96/110) to 61.8% (68/110). In addition, 4 cases were found to be misdiagnosed as benign (3 cases of non-special type of invasive breast cancer and 1 case of intraductal carcinoma in situ) and 32 cases were found to be misdiagnosed as malignant (17 cases of adenosis, 14 cases of adenosis with fibroadenoma, and 1 case of phyllodes tumor).

Multimodal ultrasound combined with artificial intelligence-based S-Detect technique to correct BI-RADS classification can improve the diagnostic efficiency for breast masses, reduce unnecessary puncture biopsies, and improve the detection rate for malignant breast masses.