Background US is clinically established for breast imaging, but its diagnostic performance depends on operator experience. Computer-assisted (real-time) image analysis may help in overcoming this limitation. Purpose To develop precise real-time-capable US-based breast tumor categorization by combining classic radiomics and autoencoder-based features from automatically localized lesions. Materials and Methods A total of 1619 B-mode US images of breast tumors were retrospectively analyzed between April 2018 and January 2024. nnU-Net was trained for lesion segmentation. Features were extracted from tumor segments, bounding boxes, and whole images using either classic radiomics, autoencoder, or both. Feature selection was performed to generate radiomics signatures, which were used to train machine learning algorithms for tumor categorization. Models were evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity and were statistically compared with histopathologically or follow-up-confirmed diagnosis. Results The model was developed on 1191 (mean age, 61 years ± 14 [SD]) female patients and externally validated on 50 (mean age, 55 years ± 15]). The development data set was divided into two parts: testing and training lesion segmentation (419 and 179 examinations) and lesion categorization (503 and 90 examinations). nnU-Net demonstrated precision and reproducibility in lesion segmentation in test set of data set 1 (median Dice score [DS]: 0.90 [IQR, 0.84-0.93]; P = .01) and data set 2 (median DS: 0.89 [IQR, 0.80-0.92]; P = .001). The best model, trained with 23 mixed features from tumor bounding boxes, achieved an AUC of 0.90 (95% CI: 0.83, 0.97), sensitivity of 81% (46 of 57; 95% CI: 70, 91), and specificity of 87% (39 of 45; 95% CI: 77, 87). No evidence of difference was found between model and human readers (AUC = 0.90 [95% CI: 0.83, 0.97] vs 0.83 [95% CI: 0.76, 0.90]; P = .55 and 0.90 vs 0.82 [95% CI: 0.75, 0.90]; P = .45) in tumor classification or between model and histopathologically or follow-up-confirmed diagnosis (AUC = 0.90 [95% CI: 0.83, 0.97] vs 1.00 [95% CI: 1.00,1.00]; P = .10). Conclusion Precise real-time US-based breast tumor categorization was developed by mixing classic radiomics and autoencoder-based features from tumor bounding boxes. ClinicalTrials.gov identifier: NCT04976257 Published under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Bahl in this issue.

中文翻译：

---

结合放射组学和自动编码器来区分美国图像上的良性和恶性乳腺肿瘤。


背景超声是临床上针对乳腺成像建立的，但其诊断性能取决于操作者的经验。计算机辅助（实时）图像分析可能有助于克服这一限制。目的 通过结合经典放射组学和自动定位病变的基于自动编码器的特征，开发精确的基于超声的实时乳腺肿瘤分类。材料和方法 回顾性分析 2018 年 4 月至 2024 年 1 月期间总共 1619 张乳腺肿瘤 B 型超声图像。 nnU-Net 接受病灶分割训练。使用经典放射组学、自动编码器或两者从肿瘤片段、边界框和整个图像中提取特征。进行特征选择以生成放射组学特征，该特征用于训练肿瘤分类的机器学习算法。使用受试者工作特征曲线下面积（AUC）、敏感性和特异性对模型进行评估，并与组织病理学或随访确诊的诊断进行统计比较。结果 该模型针对 1191 名（平均年龄，61 岁±14 [SD]）女性患者开发，并在 50 名（平均年龄，55 岁±15]）女性患者上进行了外部验证。开发数据集分为两部分：测试和训练病变分割（419和179个检查）和病变分类（503和90个检查）。 nnU-Net 在数据集 1（中位 Dice 评分 [DS]：0.90 [IQR，0.84-0.93]；P = .01）和数据集 2（中位 DS：0.89 [IQR]）的测试集中展示了病灶分割的精度和可重复性，0.80-0.92]；P = .001）。最好的模型使用来自肿瘤边界框的 23 个混合特征进行训练，其 AUC 为 0.90（95% CI：0.83，0.97），敏感性为 81%（57 中的 46；95% CI：70, 91），特异性为 87%（45 中的 39；95% CI：77, 87）。模型和人类阅读器之间没有发现差异证据（AUC = 0.90 [95% CI: 0.83, 0.97] vs 0.83 [95% CI: 0.76, 0.90]；P = .55 和 0.90 vs 0.82 [95% CI: 0.75] , 0.90]；P = .45）在肿瘤分类中或模型与组织病理学或随访确诊之间的差异（AUC = 0.90 [95% CI: 0.83, 0.97] vs 1.00 [95% CI: 1.00,1.00]；P = .10）。结论 通过混合经典放射组学和来自肿瘤边界框的基于自动编码器的特征，开发了基于 US 的精确实时乳腺肿瘤分类。 ClinicalTrials.gov 标识符：NCT04976257 根据 CC BY 4.0 许可证发布。本文提供了补充材料。另请参阅本期 Bahl 的社论。