Department of Ultrasound, Peking University Third Hospital, No. 49 North Garden Road, Haidian District, Beijing, 100191, China.
Department of General Surgery, Peking University Third Hospital, No. 49 North Garden Road, Haidian District, Beijing, 100191, China.
Eur Radiol. 2024 Nov;34(11):7080-7089. doi: 10.1007/s00330-024-10786-5. Epub 2024 May 10.
Developing a deep learning radiomics model from longitudinal breast ultrasound and sonographer's axillary ultrasound diagnosis for predicting axillary lymph node (ALN) response to neoadjuvant chemotherapy (NAC) in breast cancer.
Breast cancer patients undergoing NAC followed by surgery were recruited from three centers between November 2016 and December 2022. We collected ultrasound images for extracting tumor-derived radiomics and deep learning features, selecting quantitative features through various methods. Two machine learning models based on random forest were developed using pre-NAC and post-NAC features. A support vector machine integrated these data into a fusion model, evaluated via the area under the curve (AUC), decision curve analysis, and calibration curves. We compared the fusion model's performance against sonographer's diagnosis from pre-NAC and post-NAC axillary ultrasonography, referencing histological outcomes from sentinel lymph node biopsy or axillary lymph node dissection.
In the validation cohort, the fusion model outperformed both pre-NAC (AUC: 0.899 vs. 0.786, p < 0.001) and post-NAC models (AUC: 0.899 vs. 0.853, p = 0.014), as well as the sonographer's diagnosis of ALN status on pre-NAC and post-NAC axillary ultrasonography (AUC: 0.899 vs. 0.719, p < 0.001). Decision curve analysis revealed patient benefits from the fusion model across threshold probabilities from 0.02 to 0.98. The model also enhanced sonographer's diagnostic ability, increasing accuracy from 71.9% to 79.2%.
The deep learning radiomics model accurately predicted the ALN response to NAC in breast cancer. Furthermore, the model will assist sonographers to improve their diagnostic ability on ALN status before surgery.
Our AI model based on pre- and post-neoadjuvant chemotherapy ultrasound can accurately predict axillary lymph node metastasis and assist sonographer's axillary diagnosis.
Axillary lymph node metastasis status affects the choice of surgical treatment, and currently relies on subjective ultrasound. Our AI model outperformed sonographer's visual diagnosis on axillary ultrasound. Our deep learning radiomics model can improve sonographers' diagnosis and might assist in surgical decision-making.
从纵向乳腺超声和超声医师腋窝超声诊断中开发深度学习放射组学模型，以预测乳腺癌新辅助化疗（NAC）后腋窝淋巴结（ALN）的反应。
本研究招募了自 2016 年 11 月至 2022 年 12 月期间在三个中心接受 NAC 治疗后行手术的乳腺癌患者。我们收集了超声图像以提取肿瘤衍生的放射组学和深度学习特征，通过各种方法选择定量特征。使用基于随机森林的两种机器学习模型，基于治疗前和治疗后特征进行开发。支持向量机将这些数据集成到融合模型中，通过曲线下面积（AUC）、决策曲线分析和校准曲线进行评估。我们将融合模型的性能与治疗前和治疗后腋窝超声检查中超声医师的诊断进行了比较，参考前哨淋巴结活检或腋窝淋巴结清扫术的组织学结果。
在验证队列中，融合模型的表现优于治疗前（AUC：0.899 与 0.786，p<0.001）和治疗后模型（AUC：0.899 与 0.853，p=0.014），以及超声医师在治疗前和治疗后腋窝超声检查中对 ALN 状态的诊断（AUC：0.899 与 0.719，p<0.001）。决策曲线分析表明，该模型在阈值概率为 0.02 至 0.98 之间对患者具有益处。该模型还提高了超声医师的诊断能力，将准确率从 71.9%提高到 79.2%。
深度学习放射组学模型可准确预测乳腺癌的 ALN 对 NAC 的反应。此外，该模型将帮助超声医师提高手术前 ALN 状态的诊断能力。
我们基于新辅助化疗前后超声的人工智能模型可以准确预测腋窝淋巴结转移，并辅助超声医师的腋窝诊断。
腋窝淋巴结转移状态影响手术治疗的选择，目前依赖于主观超声。我们的 AI 模型在腋窝超声检查中优于超声医师的视觉诊断。我们的深度学习放射组学模型可以提高超声医师的诊断能力，并可能有助于手术决策。
