摘要：

准确的肿瘤亚区分割是乳腺癌异质性表征的关键, 而这种表征是了解乳腺癌化疗反应的关键。传统的阈值分割在功能性肿瘤亚区的区分上存在不足, 为此, 提出一种改进的分割(GR)方法, 该方法基于异质性影像组学特征, 包括纹理信息、强度信息、形状信息, 通过高斯混合聚类实现功能性亚区的精准分割。基于GR分割, 设计双分支双任务分类模型(DDCN), 预测新辅助化疗疗效, 利用得到的亚区来更好地提取肿瘤内部异质性的特征, 并结合文本信息来评估疗效。实验结果表明, 所提放射组学分割模型在不同功能亚区的识别上表现出色, 在剪影系数和方差比指数上都取得了良好的效果。DDCN模型融合了不同亚区的特征, 消融实验结果表明, 在受试者工作特征曲线下的面积(AUC)、准确率等指标上DDCN都取得了良好的结果。总体而言, GR在肿瘤亚区分割上比传统的阈值分割效果更优, 而DDCN模型在评估新辅助化疗疗效方面具有广泛的应用价值。

关键词: 肿瘤异质性, 聚类分割, 深度学习, 新辅助化疗, 医学影像

Abstract:

Accurate segmentation of tumor subregions is critical for characterizing the heterogeneity of breast cancer, a key factor in understanding the response to chemotherapy. Traditional threshold segmentation methods are inadequate for distinguishing functional tumor subregions. To address this, an adaptive segmentation approach called GR has been proposed. This algorithm leverages heterogeneous image omics features, including texture, strength, and shape information, using Gaussian mixture clustering to achieve precise functional subregion segmentation. Building on this, a Dual-branch Dual-task Classification Model (DDCN) was developed to predict the efficacy of neoadjuvant chemotherapy. This model utilizes the segmented subregion information to enhance feature extraction, quantifying intra-tumor heterogeneity and evaluating therapeutic outcomes through textual information. Experimental results highlight the efficacy of the radiomics-based segmentation model in identifying distinct functional subregions, outperforming traditional methods in metrics such as the silhouette coefficient and the calinski-harabasz index. Additionally, the DDCN model combines features from various subregions, achieving superior performance in metrics such as the Area Under the receiver operating characteristic Curve(AUC)and accuracy. Overall, the GR segmentation approach proves to be more effective than traditional threshold segmentation for tumor subregion analysis. The DDCN model further enhances the evaluation of neoadjuvant chemotherapy efficacy, demonstrating significant potential for broader clinical applications.

Key words: tumor heterogeneity, cluster segmentation, deep learning, neoadjuvant chemotherapy, radiomics