深度学习联合超声在乳腺BI-RADS分类中的价值应用

The Evaluating the Utility of Deep Learning Combined with Ultrasound in Breast BI-RADS Classification

摘要: 目的：评估人工智能深度学习(DL)软件优化乳腺超声BI-RADS分类的价值及联合诊断效能。方法：以病理结果作为金标准，回顾性分析237例乳腺肿块(良性153例，恶性84例)。由低年资医师按ACR BI-RADS评估肿块特征，对比DL系统诊断性能，对不一致病例分级调整后实施联合诊断。结果：超声医师诊断结果的ROC曲线下面积为0.710，AI软件为0.838；二者诊断不一致的病例进行联合诊断后，AUC提升至0.957。在形态学特征(形状、方向)的判断上，医师与AI的一致性较高；在钙化判断上一致性中等；而在内部回声、边缘、后方回声的判断上一致性较差。结论：与单一超声医师诊断相比，AI深度学习诊断系统联合超声医师优化BI-RADS分类后对乳腺癌的良恶性诊断效果更好，联合诊断能够更好地为临床决策提供更高的价值，降低不必要的活检率，避免超声诊断存在的主观性缺陷，人工智能辅助医师在诊断的过程中实现了量化及标准化，具有一定的临床推广意义。

Abstract: Objective: To evaluate the value and combined diagnostic efficacy of artificial intelligence (AI) deep learning (DL) software in optimizing BI-RADS classification for breast ultrasound. Methods: Using pathological results as the gold standard, we retrospectively analyzed 237 breast masses (153 benign, 84 malignant). Junior physicians assessed mass features according to ACR BI-RADS. DL system performance was compared, and discordant cases underwent adjusted grading for combined diagnosis. Results: The AUC for ultrasound physicians was 0.710, compared to 0.838 for the AI software. Combined diagnosis of discordant cases significantly increased the AUC to 0.957. Physician-AI agreement was high for morphological features (shape, orientation) and moderate for calcification assessment, but poor for internal echogenicity, margins, and posterior acoustic features. Conclusion: Compared to single-physician diagnosis, the AI DL system combined with physician assessment significantly improves diagnostic accuracy for breast lesion malignancy. This combined approach enhances clinical decision-making, reduces unnecessary biopsies, mitigates subjective diagnostic limitations, and introduces quantification and standardization, demonstrating significant clinical value for broader adoption.