Data Sheet 1_Intratumoral and peritumoral radiomics based on automated breast volume scanner for predicting human epidermal growth factor receptor 2 status.docx

PurposeTo develop an intratumoral and peritumoral radiomics model using Automated Breast Volume Scanner (ABVS) for noninvasive preoperative prediction of Human Epidermal Growth Factor Receptor 2 (HER2) status.
MethodsThis retrospective study analyzed 384 lesions from 379 patients with pathologically confirmed breast cancer across four hospitals. Two tasks were defined: Task 1 to distinguish HER2-negative from HER2-positive cases and Task 2 to differentiate HER2-zero from HER2-low status. For each classification task, three models were built: Model 1 included radiomics features from the tumor region alone; Model 2 included features from both the tumor region and a 5mm peritumoral region; and Model 3 incorporated features from the tumor region, the 5mm peritumoral region, and the 5-10mm peritumoral region. The performance of the model was evaluated using receiver operating characteristic (ROC) curves, with key metrics including the area under the curve (AUC), sensitivity, specificity, and accuracy.
ResultsIn the classification tasks, Model 2 demonstrated superior predictive performance across multiple datasets. For Task 1, it achieved the highest AUC (0.844), exceptional sensitivity (0.955), and satisfactory accuracy (0.787) in the validation set, and outperformed other models in the test set with an AUC of 0.749 and sensitivity of 0.885, highlighting its robustness and clinical applicability. For Task 2, Model 2 exhibited the highest AUC (0.801), sensitivity (0.862), and accuracy (0.808) in the test set, with consistent performance across the training (AUC 0.850) and validation sets (AUC 0.801). Model 3, which combines intratumoral and peritumoral features, did not demonstrate significant improvements over the intratumoral-only model in the two classification tasks. These results underscore the value of incorporating peritumoral radiomics features, particularly within a 5mm margin, to enhance predictive performance compared to intratumoral-only models.
ConclusionThe radiomics model integrating intratumoral and appropriate peritumoral features significantly outperformed the model based on intratumoral features alone. This integrated approach holds strong potential for noninvasive, preoperative prediction of HER2 status.

目的：本研究旨在利用自动乳腺容积扫描（Automated Breast Volume Scanner, ABVS）构建瘤内及瘤周放射组学模型，实现人表皮生长因子受体2（Human Epidermal Growth Factor Receptor 2, HER2）状态的无创术前预测。
方法：本项回顾性研究纳入了四家医院共379例经病理确诊的乳腺癌患者的384个病灶进行分析。设置两项分类任务：任务1用于区分HER2阴性与HER2阳性病例，任务2用于鉴别HER2零表达与HER2低表达状态。针对每项分类任务，分别构建三种模型：模型1仅纳入肿瘤区域的放射组学特征；模型2纳入肿瘤区域及5mm瘤周区域的放射组学特征；模型3整合肿瘤区域、5mm瘤周区域以及5-10mm瘤周区域的放射组学特征。采用受试者工作特征（Receiver Operating Characteristic, ROC）曲线对模型性能进行评估，核心评价指标包括曲线下面积（Area Under the Curve, AUC）、灵敏度、特异度及准确度。
结果：在两项分类任务中，模型2均展现出最优的预测性能。针对任务1，其在验证集中取得最高的曲线下面积（0.844）、优异的灵敏度（0.955）及尚可的准确度（0.787）；在测试集上同样优于其他模型，曲线下面积为0.749，灵敏度为0.885，凸显了其稳健性与临床应用价值。针对任务2，模型2在测试集上取得最高的曲线下面积（0.801）、灵敏度（0.862）及准确度（0.808），在训练集（曲线下面积0.850）与验证集（曲线下面积0.801）中亦保持一致的性能表现。整合瘤内与瘤周特征的模型3在两项分类任务中均未展现出相较于仅瘤内模型的显著性能提升。上述结果表明，相较于仅基于瘤内特征的模型，纳入瘤周放射组学特征（尤其是5mm范围内的瘤周区域）可有效提升预测性能。
结论：整合瘤内与适宜范围瘤周特征的放射组学模型，其性能显著优于仅基于瘤内特征的模型。该整合方案在HER2状态的无创术前预测中具备良好的应用前景。