Diagnostic efficacy of conventional ultrasonography combined with contrast-enhanced ultrasound versus artificial intelligence in differentiating benign and malignant breast nodules
-
摘要:
目的 评估常规超声检查(US)、超声造影(CEUS)及两者联合应用与超声人工智能(AI)辅助诊断系统对乳腺良恶性结节的诊断效能,为优化临床诊断策略提供参考。 方法 回顾性分析2024年1月-2025年3月上海交通大学医学院附属第一人民医院收治的353例经病理确诊的乳腺结节患者的影像学与临床资料。所有患者均完成US、CEUS检查及超声AI辅助诊断,采用二元logistic回归分析构建诊断模型。以病理诊断结果为金标准,绘制ROC曲线,采用灵敏度、特异度和AUC评估各方法的诊断效能。 结果 353例乳腺结节中,病理诊断为恶性136例(38.5%)、良性217例(61.5%)。US、CEUS、US+CEUS、AI及AI+临床特征5种方法诊断乳腺结节良恶性的AUC分别为0.764、0.857、0.893、0.937、0.957,其中US+CEUS诊断效能优于US或CEUS单一检查,但低于AI(均P<0.05);AI的AUC低于AI+临床特征,但高于其他3种方法(均P<0.05)。在5种方法中,CEUS的灵敏度与AI相当(均为94.9%),AI+临床特征的特异度(93.1%)、准确度(93.8%)、阳性预测值(89.6%)、阴性预测值(96.7%)最高。 结论 超声AI辅助诊断系统有助于提高乳腺结节良恶性的鉴别诊断能力,整合临床特征能进一步优化AI的诊断效能。 Abstract:Objective To evaluate the diagnostic efficacy of conventional ultrasonography (US), contrast-enhanced ultrasound (CEUS), their combination, and artificial intelligence (AI)-assisted diagnosis in differentiating benign and malignant breast nodules, so as to provide a basis for clinical diagnostic strategies. Methods The imaging and clinical data of 353 patients with pathologically confirmed breast nodules treated at Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine from Jan. 2024 to Mar. 2025 were retrospectively analyzed. All patients underwent US, CEUS, and AI-assisted ultrasound diagnostic assessments. The diagnostic models were developed using binary logistic regression. With pathology as the gold standard, receiver operating characteristic curves were generated, the sensitivity, specificity and the area under curve (AUC) were used to evaluate the diagnostic efficacy. Results Among the 353 cases, 136 were malignant (38.5%) and 217 were benign (61.5%). The AUC values of US, CEUS, their combination, AI, and AI+clinical features for differentiating benign and malignant breast nodules were 0.764, 0.857, 0.893, 0.937, and 0.957, respectively. The diagnostic efficacy of US+CEUS combination was superior to that of US or CEUS alone but inferior to that of AI (all P<0.05). The AUC of AI was significantly higher than that of the other 3 methods (US, CEUS, US+CEUS combination), but was significantly lower than that of the AI+clinical features (all P<0.05). The sensitivity of CEUS was comparable to that of AI (both 94.9%). Moreover, AI+clinical features achieved the highest specificity (93.1%), accuracy (93.8%), positive predictive value (89.6%), and negative predictive value (96.7%). Conclusion The AI-assisted ultrasound diagnostic system can improve the differential diagnosis of benign and malignant breast nodules, and integrating clinical features into AI can optimize the diagnostic efficacy of AI. -
乳腺癌位居全球女性恶性肿瘤发病率首位[1],早期精准诊断对改善预后至关重要。超声检查因无辐射、便捷、可重复及实时成像等优势,已成为我国乳腺病变筛查的首选影像学方法[2]。美国放射学会发布的第5版乳腺影像报告和数据系统(Breast Imaging Reporting and Data System,BI-RADS)将乳腺结节分为6类[3],然而在常规超声检查(conventional ultrasonography,US)图像中,不同类别结节的良恶性特征常存在交叉重叠,鉴别诊断难度较大。此外,US对结节声像特征的判读高度依赖医师经验,主观性强,诊断一致性不足,亟需引入更客观的评价手段以降低操作者间差异[4-5]。超声造影(contrast-enhanced ultrasound,CEUS)可实时动态显示病灶微血管灌注,清晰勾勒感兴趣区域(region of interest,ROI)的血流分布,有助于提升微小病灶的检出能力,然而,良恶性结节在CEUS中仍存在重叠表现,限制了其临床应用的广泛性[6]。穿刺活检病理诊断虽是诊断乳腺结节的金标准,但其为有创性检查,并且取材偏差等问题可能影响诊断准确性[7]。
近年来,人工智能(artificial intelligence,AI)技术在乳腺结节诊断中的应用日益成熟[5, 8]。AI可自动定位乳腺结节,提取其形态学与边界等关键特征,实现自动化分析与诊断,为良恶性鉴别提供客观、量化的依据,辅助超声医师提升诊断准确性和一致性,并在早期病变识别方面展现出重要潜力[4, 9]。本研究系统评估US和CEUS对乳腺良恶性结节的诊断效能,同时分析超声AI辅助诊断系统的鉴别性能,比较US+CEUS联合诊断与超声AI辅助诊断的差异,以期为临床优化乳腺结节的诊断策略提供参考。
1 资料和方法
1.1 研究对象
回顾性纳入2024年1月至2025年3月于上海交通大学医学院附属第一人民医院经穿刺活检或手术病理确诊的353例18岁以上女性乳腺结节患者。纳入标准:(1)完成US、CEUS检查及超声AI辅助诊断;(2)结节在图像中完整显示,未超出边界;(3)超声图像完整,且病理结果明确。排除标准:(1)妊娠期或哺乳期女性;(2)曾接受乳腺假体植入术或新辅助化疗者;(3)合并严重全身性疾病、不能耐受检查或依从性差者;(4)对超声造影剂过敏者;(5)存在多发相邻结节影响ROI准确勾画者。本研究经我院医学伦理委员会审批。
1.2 仪器与方法
US、CEUS检查均使用美国GE HealthCare公司生产的LOGIQ E9彩色多普勒超声诊断仪。US使用ML6-15线阵探头,频率6~15 MHz;CEUS使用9L线阵探头,频率2~8 MHz,造影剂采用意大利Bracco公司生产的注射用六氟化硫微泡(SonoVue)。超声AI辅助诊断使用脉得智能科技(无锡)有限公司研发的乳腺结节超声辅助诊断系统。
1.2.1 US检查
患者取仰卧位,双手上举,充分暴露双侧乳房及腋窝,对双侧乳房进行全方位扫查。确定目标病灶后,记录病灶超声图像特征信息,包括病灶的位置、大小、形态、边界、回声、钙化、血流、腋下淋巴结等。采用第5版BI-RADS标准[3]对结节进行综合性评估。
1.2.2 CEUS检查
患者体位同US检查。进行US扫查明确目标结节后,选清晰显示目标结节及周围乳腺腺体的最佳切面,切换至CEUS模式,调节增益、深度及聚焦等至适宜范围,嘱患者保持平静呼吸。用5 mL生理盐水稀释造影剂并震荡混匀后,经肘正中静脉团注4.8 mL造影剂悬浮液,随即快速推注5 mL生理盐水,动态采集结节CEUS图像并连续动态存储至少3 min[10]。CEUS评估指标包括:(1)增强强度(与周围正常组织在峰值时比较,高增强或低增强);(2)增强均匀性(均匀或不均匀);(3)增强范围(扩大或未扩大);(4)增强边缘(清晰或不清晰);(5)是否存在滋养血管(有或无);(6)是否存在周边放射状增强(有或无)。采用第5版BI-RADS标准[3]对结节进行综合性评估。
1.2.3 超声AI辅助诊断
将目标结节US的最大切面图像导入基于深度学习框架的超声AI辅助诊断系统,该系统运行流程分两个阶段。(1)病灶识别:通过卷积神经网络结合多尺度特征提取与注意力机制,定位并自动分割乳腺灰阶图像中的可疑病灶,实现精准识别。(2)分类判别:借助多层级特征融合网络整合影像特征,提取反映病灶形态、回声、钙化等图像特征信息,通过分类网络判断良恶性。为提升模型性能,训练时引入图像旋转、噪声扰动、亮度归一化等数据增强策略及领域自适应机制,适配不同设备与图像质量;同时采用迁移学习与自蒸馏优化,缩短训练时间并强化特征提炼,且模型可输出预测可靠性以保障临床应用安全。系统设定0.5作为量化参数的判定阈值,数值≥0.5表示具备相应恶性特征。根据以上量化指标,AI软件自动将结节分类为良性或恶性,并输出相应概率。
1.3 统计学处理
采用SPSS 26.0软件进行数据分析。计数资料以例数和百分数表示,组间比较采用χ2检验;计量资料以 x±s表示,组间比较采用独立样本t检验。采用二元logistic回归分析构建诊断模型。以病理结果为金标准,绘制各种方法鉴别乳腺结节良恶性的ROC曲线,计算AUC、灵敏度、特异度、准确度,采用DeLong检验比较不同方法AUC的差异。检验水准(α)为0.05。
2 结果
2.1 病理检查结果
2024年1月至2025年3月在我院行US检查提示乳腺结节、建议行CEUS者共831例,其中507例进行了CEUS及超声AI辅助诊断,排除未在我院进行穿刺或手术者132例、病理结果不明确者6例、有既往手术或放化疗病史者16例,共纳入353例患者353个乳腺结节。患者年龄22~89(54.0±7.5)岁。病灶短径0.20~5.20(1.04±0.67)cm,长径0.30~8.10(1.15±0.83)cm。术后病理证实恶性结节136例(38.5%),包括非特殊浸润性乳腺癌121例、导管内原位癌7例、浸润性小叶癌5例、包裹性乳头状癌2例和浸润性微乳头状癌1例;良性病灶217例(61.5%),包括纤维腺瘤83例、硬化性腺病57例、导管内乳头状瘤47例、增生性病变18例、乳腺炎8例、导管扩张4例。统计分析显示,恶性结节组患者年龄大于良性结节组,有家族史和绝经的比例高于良性结节组,差异均有统计学意义(均P<0.01,表 1)。
表 1 病理诊断为恶性和良性乳腺结节患者的基线资料及US特征Table 1 Baseline data and US characteristics of patients with pathologically diagnosed malignant and benign breast nodulesFeature Malignant nodule N = 136 Benign nodule N = 217 Statistic P value Age/year, x±s 57.71±13.52 50.25±13.57 t = 5.032 0.001 Family history, n (%) χ2 = 9.711 0.002 Yes 22 (16.18) 13 (5.99) No 114 (83.82) 204 (94.01) Menopause, n (%) χ2 = 11.811 0.001 Yes 90 (66.18) 103 (47.47) No 46 (33.82) 114 (52.53) Hypoechoic, n (%) χ2 = 3.236 0.072 Yes 133 (97.79) 200 (92.17) No 3 (2.21) 17 (7.83) Indistinct margin, n (%) χ2 = 40.05 <0.001 Yes 106 (77.94) 94 (43.32) No 30 (22.06) 123 (56.68) Irregular shape, n (%) χ2 = 26.544 <0.001 Yes 125 (91.91) 147 (67.74) No 11 (8.09) 70 (32.26) Calcification, n (%) χ2 = 18.681 <0.001 Yes 52 (38.24) 38 (17.51) No 84 (61.76) 179 (82.49) Blood flow, n (%) χ2 = 21.361 <0.001 Yes 93 (68.38) 93 (42.86) No 43 (31.62) 124 (57.14) US: Conventional ultrasonography. 2.2 乳腺良恶性结节US及CEUS特征分析
乳腺恶性和良性结节的典型US及CEUS图像见图 1、图 2。在US特征中,恶性结节表现为边界不清晰、形态不规则、有钙化、有血流的比例均高于良性结节(均P<0.001,表 1),CEUS特征中,恶性结节呈现高增强、不均匀增强、增强范围扩大、增强边缘不清晰、有滋养血管及周边有放射状增强的比例均高于良性结节(均P<0.01,表 2)。
图 1 1例61岁女性患者乳腺恶性结节US、CEUS图像及AI辅助诊断结果(病理证实为浸润性乳腺癌)Fig. 1 US and CEUS images and AI-assisted diagnosis results of a malignant breast nodule in a 61-year-old female patient (invasive breast cancer confirmed by pathology)A: US revealed a hypoechoic mass at the 10 o'clock position of the right breast, measuring 28 mm×19 mm, with indistinct and spiculated margin, irregular shape, posterior acoustic shadowing; B: CEUS demonstrated heterogeneous hyperenhancement, the lesion size post-enhancement appeared larger compared to US, with indistinct margin after enhancement, the presence of nutrient vessels, and peripheral radial enhancement, and was diagnosed as "suspicious for malignancy"; C, D: The breast nodule was identified and assessed by AI software, which estimated a 93% probability of malignancy. US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence.
图 2 1例38岁女性患者乳腺良性结节US、CEUS图像及AI辅助诊断结果(病理证实为纤维腺瘤)Fig. 2 US and CEUS images and AI-assisted diagnosis results of a benign breast nodule in a 38-year-old female patient (fibroadenoma confirmed by pathology)A: US revealed a hypoechoic nodule at the 10 o'clock position of the right breast, measuring 16 mm×11 mm, with indistinct margin, a microlobulated shape, heterogeneous internal echotexture, and punctate internal blood flow signals, with low-velocity and low-resistance spectrum detected; B: CEUS demonstrated homogeneous hyperenhancement, the lesion size post-enhancement appeared larger compared to US, with distinct margin after enhancement, the presence of nutrient vessels, and no peripheral radial enhancement, and was diagnosed as "suspicious for malignancy"; C, D: The breast nodule was identified and assessed by AI software, which estimated a 78% probability of benign tumor. US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence.表 2 病理诊断为恶性和良性乳腺结节患者的CEUS特征Table 2 CEUS characteristics of patients with pathologically diagnosed malignant and benign berast nodulesn (%) Feature Malignant nodule N =136 Benign nodule N =217 χ2 value P value Hyperenhancement 118 (86.76) 159 (73.27) 25.045 <0.001 Heterogeneous enhancement 100 (73.53) 98 (45.16) 27.317 <0.001 Expanded enhancement range 85 (62.50) 60 (27.65) 41.952 <0.001 Indistinct enhancement margin 80 (58.82) 93 (42.86) 8.528 0.003 Presence of nutrient vessels 132 (97.06) 125 (57.60) 65.732 <0.001 Peripheral radial enhancement 22 (16.18) 3 (1.38) 25.603 <0.001 CEUS: Contrast-enhanced ultrasound. 2.3 不同方法诊断性能分析
多因素logistic回归分析结果显示,US特征中边界、钙化、血流为独立预测因素(均P<0.05),CEUS特征中增强均匀性、增强范围、是否存在滋养血管、是否存在周边放射状增强为独立预测因素(均P<0.05),而在AI+临床特征中AI诊断结果与年龄为独立预测因素(均P<0.05),见表 3。基于logistic回归分析结果构建4种诊断模型:模型1整合US特征,模型2整合CEUS特征,模型3整合US特征及CEUS特征,模型4整合AI诊断结果及临床特征。
表 3 US特征、CEUS特征、AI+临床特征诊断乳腺结节良恶性的logistic回归分析结果Table 3 Results of logistic regression analysis for differentiating benign and malignant breast nodules based on US features, CEUS features, and AI+clinical featuresVariable b SE Wald χ2 value P value OR (95%CI) US Indistinct margin 1.150 0.281 16.721 <0.001 3.160 (1.820, 5.484) Irregular shape 0.708 0.385 3.386 0.066 2.031 (0.955, 4.318) Calcification 0.913 0.279 10.713 0.001 2.492 (1.442, 4.306) Blood flow 1.039 0.252 17.044 <0.001 2.826 (1.726, 4.628) Hypoechoic 0.568 0.652 0.758 0.384 1.764 (0.491, 6.331) Constant -5.098 0.768 44.107 <0.001 0.006 CEUS Heterogeneous enhancement 1.751 0.305 32.863 <0.001 5.759 (3.165, 10.478) Expanded enhancement range 1.336 0.307 18.924 <0.001 3.803 (2.083, 6.941) Presence of nutrient vessels 3.419 0.603 32.124 <0.001 30.553 (9.365, 99.679) Peripheral radial enhancement 2.252 0.687 10.760 0.001 9.507 (2.476, 36.513) Hyperenhancement -0.464 0.434 1.148 0.284 0.628 (0.854, 2.697) Indistinct enhancement margin 0.417 0.293 2.025 0.155 1.518 (0.269, 1.470) Constant -11.759 1.662 50.052 <0.001 0.000 AI+clinical features AI 5.593 0.520 115.718 <0.001 268.578 (96.938, 744.124) Age 0.097 0.033 8.532 0.003 1.101 (1.032, 1.175) Menopause 1.473 0.825 3.187 0.074 4.363 (0.866, 21.990) Family history -0.701 0.797 0.775 0.379 0.496 (0.104, 2.364) Constant -9.422 3.266 8.324 0.004 0.000 US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; b: Regression coefficient; SE: Standard error; OR: Odds ratio; 95%CI: 95% confidence interval. 以术后病理结果为金标准,绘制US、CEUS、US+CEUS、AI及AI+临床特征5种诊断方法的ROC曲线(图 3),各方法的诊断效能指标见表 4,其AUC分别为0.764、0.857、0.893、0.937、0.957(均P<0.05)。AI+临床特征的AUC高于US、CEUS、US+CEUS和AI,差异均有统计学意义(Z=6.546,P<0.001;Z=3.324,P<0.001;Z=2.916,P=0.004;Z=2.365,P=0.018);AI的AUC低于AI+临床特征,但高于US、CEUS、US+CEUS(Z=2.365,P=0.018;Z=6.132,P<0.001;Z=3.406,P<0.001;Z=2.020,P=0.043)。CEUS、AI及AI+临床特征的灵敏度相当;AI+临床特征在特异度、准确度、阳性预测值和阴性预测值上均为最高,AI次之。US+CEUS联合诊断的AUC优于US和CEUS单独应用(Z=5.762,P<0.001;Z=3.393,P<0.001),特异度、准确度及阳性预测值也高于两者单独应用,但仍低于AI。
图 3 US、CEUS、US+CEUS、AI、AI+临床特征鉴别诊断乳腺良恶性结节的ROC曲线Fig. 3 ROC curves for differential diagnosis of benign and malignant breast nodules using US, CEUS, US+CEUS, AI, and AI+clinical featuresUS: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; ROC: Receiver operating characteristic.表 4 不同诊断方式的诊断效能比较Table 4 Comparison of diagnostic efficacy among different diagnostic methodsMethod AUC Sensitivity/% (n/N) Specificity/% (n/N) Accuracy/% (n/N) Positive predictive value/% (n/N) Negative predictive value/% (n/N) US 0.764 65.4 (89/136) 80.2 (174/217) 74.5 (263/353) 67.4 (89/132) 78.7 (174/221) CEUS 0.857 94.9 (129/136) 63.1 (137/217) 75.4 (266/353) 61.7 (129/209) 95.1 (137/144) US+CEUS 0.893 80.9 (110/136) 83.9 (182/217) 82.7 (292/353) 75.9 (110/145) 87.5 (182/208) AI 0.937 94.9 (129/136) 92.6 (201/217) 93.5 (330/353) 89.0 (129/145) 96.6 (201/208) AI+clinical features 0.957 94.9 (129/136) 93.1 (202/217) 93.8 (331/353) 89.6 (129/144) 96.7 (202/209) US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; AUC: Area under curve. 3 讨论
早期精准诊断对乳腺癌患者治疗决策的制定及预后改善至关重要[11]。本研究比较了US、CEUS、US+CEUS、AI及AI+临床特征在乳腺良恶性结节鉴别中的诊断效能。结果显示,AI +临床特征的诊断效能最优(AUC为0.957),高于单独应用AI(AUC为0.937)及其他3种方法;US+CEUS的诊断效能(AUC为0.893)虽优于单一检查方法,但仍低于AI;CEUS单独应用的诊断能力亦优于US。上述结果提示,整合临床特征可进一步提升AI的诊断准确率,AI和多模态影像结合在乳腺结节诊断中具有较大潜力。
US作为乳腺结节筛查的主要方法,在本研究中诊断效能相对有限(AUC为0.764,灵敏度为65.4%,特异度为80.2%)。其局限性主要源于对操作者经验的依赖较强,诊断主观性较高。尽管美国放射学会第5版BI-RADS对结节进行了标准化分类[3],但US中不同类别结节的良恶性声像特征仍存在较多重叠,难以实现精准鉴别,这也影响了US诊断的一致性与准确性。
CEUS能实时动态显示病灶微血流灌注情况,本研究中其AUC为0.857,灵敏度高达94.9%,显示出其在识别恶性结节方面的优势。该优势源于恶性结节多具有更丰富且结构异常的血管分布,CEUS可清晰显示ROI的血流分布,从而提高了对恶性结节的检出能力[12]。然而,其特异度仅为63.1%,说明仍有部分良性结节被误判为恶性,这可能与良恶性结节的CEUS特征存在重叠有关,限制了其特异度的进一步提升。本研究中共有87例CEUS误诊病例,其中80例良性结节被误判为恶性,病理类型包括乳腺炎、导管内乳头状瘤、纤维腺瘤及硬化性腺病。误诊原因可能包括:炎症导致新生血管丰富,易被判断为恶性;部分导管内乳头状瘤与早期导管内原位癌的CEUS表现相似;部分纤维腺瘤和硬化性腺病可表现为不均匀高增强、增强后边界不清等类似恶性的特征[10, 13]。另有7例恶性结节被误判为良性,主要为导管原位癌和部分浸润性癌,其原因可能与病灶较小或新生血管生成活性低导致的血流灌注不明显有关。结合术后免疫组织化学结果分析,这些误判病例中人表皮生长因子受体2及Ki-67表达阴性或处于极低水平,提示其新生血管生成及癌细胞增殖活性较低,血管结构改变不明显,因此在CEUS中表现出类似良性的增强特征[14]。
US与CEUS联合应用的AUC达到0.893,较两者单独应用时均有提升,其特异度、准确度、阳性预测值亦有明显提高。US能提供结节的位置、大小、形态等基础信息,而CEUS补充了微血流灌注特征,两者结合实现了结构信息与功能信息的互补,有效弥补了单一成像模式的局限,从而提升了诊断效能。本研究结果与余丽平等[15]的报道一致,进一步验证了US与CEUS联合诊断的临床价值。尽管如此,该联合方法的AUC仍低于AI,提示其在鉴别诊断准确性方面仍有提升空间。误诊病例分析显示,部分良性结节(如伴炎症反应者)表现出类似恶性的影像特征,而部分恶性结节(如导管内癌)则因体积小、边界清晰而呈现良性特征。对于此类影像学表现不典型的病变,仍需结合穿刺活检以明确诊断。
AI在本研究中表现出较高的诊断效能,AUC达0.937,灵敏度为94.9%,特异度为92.6%,准确度为93.5%。本研究使用的超声AI辅助诊断系统能够快速、精准锁定乳腺结节,自动提取其形态、边界、回声等多维度特征并进行客观、量化分析,有效降低了人为判断的主观性和经验差异。其内置的量化参数以0.5作为判别阈值,有助于更精确地界定结节特征,从而提升诊断的一致性与可靠性。本研究中AI的AUC(0.937)较高的可能原因包括以下几个方面:首先,AI模型训练基于来自十余家医院的多中心影像数据,具备良好的泛化基础;第二,为增强数据的多样性与代表性,训练集涵盖多品牌超声设备及不同探头参数,并纳入不同年龄段与病理类型的结节图像;第三,在图像预处理阶段采用自研算法进行质量筛查与伪影检测,排除噪声干扰严重或时间-强度曲线异常的图像,确保输入数据质量;第四,标注过程采用“AI预标注-医师校准-专家复核”三级协同机制,由AI完成初步识别与标注,经临床医师修正和高年资医师审核后,最终由乳腺影像专家组复核确认,形成高质量的标准化标注数据集,为模型训练与验证提供了可靠基础。
在本研究AI误诊的23例乳腺结节中,16例良性结节被误判为恶性,主要原因包括结节体积较大影响识别,部分良性病变(如硬化性腺病)在超声图像中呈现类似恶性的特征,以及炎症病灶边界模糊形成假浸润表现。另有7例恶性结节被误判为良性,可能与病灶较小或其量化特征与良性结节相似有关,此发现与Xiao等[16]的研究结论一致。针对上述误诊情况,建议临床实践中采取以下策略:对AI提示恶性但CEUS未见典型恶性特征的结节,可建议进行3~6个月的短期超声随访;而对于AI提示良性但具有乳腺癌家族史等高危因素的患者,应考虑适当放宽穿刺活检指征,以降低漏诊风险。AI在乳腺良恶性结节鉴别中表现出高诊断效能,有望成为一种有效的术前无创诊断手段,对于US及CEUS难以明确性质的结节,AI可提供更客观、准确的参考依据,有助于减少不必要的穿刺活检,减轻患者痛苦与经济负担。
本研究结果显示,结合临床特征的AI模型诊断效能最优,AUC达0.957,灵敏度为94.9%,特异度为93.1%,准确度为93.8%。与单纯基于超声图像的AI辅助诊断相比,其诊断效能虽未出现大幅度跃升,但这一微小提升在临床实践中仍具有重要意义,有助于减少假阳性导致的过度活检或假阴性引发的漏诊风险。该结果不仅验证了临床信息对AI诊断系统的补充价值,也为构建多维度信息整合的乳腺结节诊断模式提供了实践依据。
本研究存在一定局限性。首先,回顾性设计及单中心样本可能存在选择偏倚,样本量有限,未来计划联合多家三甲医院,纳入1 000例以上病例开展前瞻性研究,扩大样本地域与人群多样性,验证超声AI辅助诊断系统的通用性。其次,当前AI模型仅基于静态图像分析,后续可收集CEUS动态血流灌注序列数据,开发基于时空特征的AI模型,以提升对微小和不典型病灶的识别能力。本研究所使用的超声图像均来自单一品牌及型号设备,且由经验丰富的超声医师采集,模型在其他设备或常规操作条件下的适用性仍需验证,后续可进一步整合多品牌、多中心图像数据,优化模型对不同图像质量的适应性,并探索AI与钼靶、MRI等其他影像技术联合应用的诊断价值,以构建更全面、精准的乳腺结节诊断体系。
综上所述,超声AI辅助诊断系统与CEUS均是鉴别乳腺结节良恶性的有效方法,US联合CEUS可优势互补,较单一方法诊断效能更高;AI的诊断效能优于US联合CEUS,在减少人为误差、提升图像分析一致性、实现自动化测量、加速诊断流程方面具有良好的应用前景;整合临床信息构建的“影像+临床”多模态诊断模型能够通过多维度信息弥补单一影像诊断的局限性,进一步提高诊断效能。
-
图 1 1例61岁女性患者乳腺恶性结节US、CEUS图像及AI辅助诊断结果(病理证实为浸润性乳腺癌)
Fig. 1 US and CEUS images and AI-assisted diagnosis results of a malignant breast nodule in a 61-year-old female patient (invasive breast cancer confirmed by pathology)
A: US revealed a hypoechoic mass at the 10 o'clock position of the right breast, measuring 28 mm×19 mm, with indistinct and spiculated margin, irregular shape, posterior acoustic shadowing; B: CEUS demonstrated heterogeneous hyperenhancement, the lesion size post-enhancement appeared larger compared to US, with indistinct margin after enhancement, the presence of nutrient vessels, and peripheral radial enhancement, and was diagnosed as "suspicious for malignancy"; C, D: The breast nodule was identified and assessed by AI software, which estimated a 93% probability of malignancy. US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence.
图 2 1例38岁女性患者乳腺良性结节US、CEUS图像及AI辅助诊断结果(病理证实为纤维腺瘤)
Fig. 2 US and CEUS images and AI-assisted diagnosis results of a benign breast nodule in a 38-year-old female patient (fibroadenoma confirmed by pathology)
A: US revealed a hypoechoic nodule at the 10 o'clock position of the right breast, measuring 16 mm×11 mm, with indistinct margin, a microlobulated shape, heterogeneous internal echotexture, and punctate internal blood flow signals, with low-velocity and low-resistance spectrum detected; B: CEUS demonstrated homogeneous hyperenhancement, the lesion size post-enhancement appeared larger compared to US, with distinct margin after enhancement, the presence of nutrient vessels, and no peripheral radial enhancement, and was diagnosed as "suspicious for malignancy"; C, D: The breast nodule was identified and assessed by AI software, which estimated a 78% probability of benign tumor. US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence.
图 3 US、CEUS、US+CEUS、AI、AI+临床特征鉴别诊断乳腺良恶性结节的ROC曲线
Fig. 3 ROC curves for differential diagnosis of benign and malignant breast nodules using US, CEUS, US+CEUS, AI, and AI+clinical features
US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; ROC: Receiver operating characteristic.
表 1 病理诊断为恶性和良性乳腺结节患者的基线资料及US特征
Table 1 Baseline data and US characteristics of patients with pathologically diagnosed malignant and benign breast nodules
Feature Malignant nodule N = 136 Benign nodule N = 217 Statistic P value Age/year, x±s 57.71±13.52 50.25±13.57 t = 5.032 0.001 Family history, n (%) χ2 = 9.711 0.002 Yes 22 (16.18) 13 (5.99) No 114 (83.82) 204 (94.01) Menopause, n (%) χ2 = 11.811 0.001 Yes 90 (66.18) 103 (47.47) No 46 (33.82) 114 (52.53) Hypoechoic, n (%) χ2 = 3.236 0.072 Yes 133 (97.79) 200 (92.17) No 3 (2.21) 17 (7.83) Indistinct margin, n (%) χ2 = 40.05 <0.001 Yes 106 (77.94) 94 (43.32) No 30 (22.06) 123 (56.68) Irregular shape, n (%) χ2 = 26.544 <0.001 Yes 125 (91.91) 147 (67.74) No 11 (8.09) 70 (32.26) Calcification, n (%) χ2 = 18.681 <0.001 Yes 52 (38.24) 38 (17.51) No 84 (61.76) 179 (82.49) Blood flow, n (%) χ2 = 21.361 <0.001 Yes 93 (68.38) 93 (42.86) No 43 (31.62) 124 (57.14) US: Conventional ultrasonography. 表 2 病理诊断为恶性和良性乳腺结节患者的CEUS特征
Table 2 CEUS characteristics of patients with pathologically diagnosed malignant and benign berast nodules
n (%) Feature Malignant nodule N =136 Benign nodule N =217 χ2 value P value Hyperenhancement 118 (86.76) 159 (73.27) 25.045 <0.001 Heterogeneous enhancement 100 (73.53) 98 (45.16) 27.317 <0.001 Expanded enhancement range 85 (62.50) 60 (27.65) 41.952 <0.001 Indistinct enhancement margin 80 (58.82) 93 (42.86) 8.528 0.003 Presence of nutrient vessels 132 (97.06) 125 (57.60) 65.732 <0.001 Peripheral radial enhancement 22 (16.18) 3 (1.38) 25.603 <0.001 CEUS: Contrast-enhanced ultrasound. 表 3 US特征、CEUS特征、AI+临床特征诊断乳腺结节良恶性的logistic回归分析结果
Table 3 Results of logistic regression analysis for differentiating benign and malignant breast nodules based on US features, CEUS features, and AI+clinical features
Variable b SE Wald χ2 value P value OR (95%CI) US Indistinct margin 1.150 0.281 16.721 <0.001 3.160 (1.820, 5.484) Irregular shape 0.708 0.385 3.386 0.066 2.031 (0.955, 4.318) Calcification 0.913 0.279 10.713 0.001 2.492 (1.442, 4.306) Blood flow 1.039 0.252 17.044 <0.001 2.826 (1.726, 4.628) Hypoechoic 0.568 0.652 0.758 0.384 1.764 (0.491, 6.331) Constant -5.098 0.768 44.107 <0.001 0.006 CEUS Heterogeneous enhancement 1.751 0.305 32.863 <0.001 5.759 (3.165, 10.478) Expanded enhancement range 1.336 0.307 18.924 <0.001 3.803 (2.083, 6.941) Presence of nutrient vessels 3.419 0.603 32.124 <0.001 30.553 (9.365, 99.679) Peripheral radial enhancement 2.252 0.687 10.760 0.001 9.507 (2.476, 36.513) Hyperenhancement -0.464 0.434 1.148 0.284 0.628 (0.854, 2.697) Indistinct enhancement margin 0.417 0.293 2.025 0.155 1.518 (0.269, 1.470) Constant -11.759 1.662 50.052 <0.001 0.000 AI+clinical features AI 5.593 0.520 115.718 <0.001 268.578 (96.938, 744.124) Age 0.097 0.033 8.532 0.003 1.101 (1.032, 1.175) Menopause 1.473 0.825 3.187 0.074 4.363 (0.866, 21.990) Family history -0.701 0.797 0.775 0.379 0.496 (0.104, 2.364) Constant -9.422 3.266 8.324 0.004 0.000 US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; b: Regression coefficient; SE: Standard error; OR: Odds ratio; 95%CI: 95% confidence interval. 表 4 不同诊断方式的诊断效能比较
Table 4 Comparison of diagnostic efficacy among different diagnostic methods
Method AUC Sensitivity/% (n/N) Specificity/% (n/N) Accuracy/% (n/N) Positive predictive value/% (n/N) Negative predictive value/% (n/N) US 0.764 65.4 (89/136) 80.2 (174/217) 74.5 (263/353) 67.4 (89/132) 78.7 (174/221) CEUS 0.857 94.9 (129/136) 63.1 (137/217) 75.4 (266/353) 61.7 (129/209) 95.1 (137/144) US+CEUS 0.893 80.9 (110/136) 83.9 (182/217) 82.7 (292/353) 75.9 (110/145) 87.5 (182/208) AI 0.937 94.9 (129/136) 92.6 (201/217) 93.5 (330/353) 89.0 (129/145) 96.6 (201/208) AI+clinical features 0.957 94.9 (129/136) 93.1 (202/217) 93.8 (331/353) 89.6 (129/144) 96.7 (202/209) US: Conventional ultrasonography; CEUS: Contrast-enhanced ultrasound; AI: Artificial intelligence; AUC: Area under curve. -
[1] 莫淼,王泽洲,郑莹,等.2022年全球及中国乳腺癌流行病学特征分析[J].海军军医大学学报,2025,46(4):497-503. DOI: 10.16781/j.CN31-2187/R.20240649. [2] GUO R, LU G, QIN B, et al. Ultrasound imaging technologies for breast cancer detection and management: a review[J]. Ultrasound Med Biol, 2018, 44(1): 37-70. DOI: 10.1016/j.ultrasmedbio.2017.09.012. [3] D'ORSI C J, MORRIS E A, D'ORSI C, et al. ACR BI-RADS atlas: Breast Imaging Reporting and Data System[M]. 5th ed. Reston, Virginia: American College of Radiology, 2013. [4] 陈蕊, 吴墅, 郭佳, 等. 超声人工智能辅助诊断系统在最大径≤ 2 cm的BI-RADS 4类乳腺结节诊断中的应用价值[J]. 海军军医大学学报, 2024, 45(5): 592-598. DOI: 10.16781/j.CN31-2187/R.20230009. [5] 吴墅, 赵佳琦, 陈蕊. 人工智能自动检测系统辅助超声诊断乳腺结节的临床应用[J]. 同济大学学报(医学版), 2022, 43(4): 509-514. DOI: 10.12289/j.issn.1008-0392.22165. [6] 李晓宇, 刘利平, 辛雨薇, 等. 人工智能软件联合超声造影鉴别诊断甲状腺良、恶性结节[J]. 中国医学影像学杂志, 2023, 31(3): 226-230. DOI: 10.3969/j.issn.1005-5185.2023.03.007. [7] HE P, CUI L G, CHEN W, et al. Subcategorization of ultrasonographic BI-RADS category 4: assessment of diagnostic accuracy in diagnosing breast lesions and influence of clinical factors on positive predictive value[J]. Ultrasound Med Biol, 2019, 45(5): 1253-1258. DOI: 10.1016/j.ultrasmedbio.2018.12.008. [8] 周杨, 周秀, 周志惠, 等. 人工智能联合超声检查在乳腺结节良恶性鉴别诊断中的应用效能[J]. 医学影像学杂志, 2025, 35(6): 169-171, 175. DOI: 10.20258/j.cnki.1006-9011.2025.06.037. [9] 罗雪, 魏红梅, 高国强, 等. 超声人工智能联合剪切波弹性成像用于乳腺结节良恶性鉴别诊断临床研究[J]. 影像科学与光化学, 2024, 42(5): 429-435. DOI: 10.7517/issn.1674-0475.2024.05.04. [10] 梁汝娜, 吴荣鹏, 连俊, 等. 超声造影定性特征与定量参数鉴别诊断BI-RADS 4类结节良恶性的对比研究[J]. 临床超声医学杂志, 2025, 27(6): 471-477. DOI: 10.16245/j.cnki.issn1008-6978.2025.06.001. [11] 中国研究型医院学会乳腺专业委员会中国女性乳腺癌筛查指南制定专家组. 中国女性乳腺癌筛查指南(2022年版)[J]. 中国研究型医院, 2022, 9(2): 6-13. DOI: 10.19450/j.cnki.jcrh.2022.02.003. [12] 李晨, 谷立爽, 刘媛, 等. 超声造影对乳腺癌新辅助化疗疗效评估的系列测量分析[J]. 国际医学放射学杂志, 2025, 48(4): 397-402. DOI: 10.19300/j.2025.L22088. [13] 吴娱, 戴海霞, 徐晓红. 二维超声联合超声造影在不同病理亚型乳腺叶状肿瘤鉴别诊断中的应用价值[J]. 中国超声医学杂志, 2024, 40(6): 656-660. DOI: 10.3969/j.issn.1002-0101.2024.06.016. [14] 张宇, 杨乐, 徐鑫. 超声造影TIC定量参数与乳腺癌HER-2、Ki-67表达状态的关系[J]. 分子诊断与治疗杂志, 2025, 17(6): 1058-1061. DOI: 10.19930/j.cnki.jmdt.2025.06.022. [15] 余丽平, 康春梅, 何志安, 等. 超声造影联合常规超声与磁共振成像对乳腺恶性非肿块病变的诊断价值比较研究[J]. 中国医学装备, 2025, 22(5): 72-77. DOI: 10.3969/j.issn.1672-8270.2025.05.015. [16] XIAO M, ZHAO C, ZHU Q, et al. An investigation of the classification accuracy of a deep learning framework-based computer-aided diagnosis system in different pathological types of breast lesions[J]. J Thorac Dis, 2019, 11(12): 5023-5031. DOI: 10.21037/jtd.2019.12.10.
下载: