Efficient and cost-effective breast cancer risk stratification using whole slide histopathology images

使用全载玻片组织病理学进行有效且具有成本效益的乳腺癌风险分层 图像 乳腺癌预后高度依赖于受体状态，因为最佳治疗取决于 是否存在雌激素、孕酮或HER-2/neu受体的过表达。 为了防止过度治疗化疗患者，关键是量化 雌激素受体（ER）阳性（ER+）、HER 2阴性（HER 2-）乳腺癌复发。一 满足这种需要的常用评估方法是Oncotype DX（ODX）复发评分。 不幸的是，ODX和类似的基因测定是昂贵的，耗时的，并且组织损伤。 破坏性的作为替代方案，我们建议使用常规方法估计ODX复发评分， 普遍存在的和廉价的苏木精和伊红（H&E）染色的活组织检查。还有其他 努力预测来自H&E的ODX复发风险。这些自动化方法检测组织学 基元（例如，细胞核）中，通常在特定的，也是自动检测的，解剖区域（例如， 导管、小管、管腔、上皮和基质）。将风险分为两种或三种 分类，通常将两个类别合并为一个。这些模型的性能是 很有前途但还是很谦虚提高模型性能的一种方法是训练 更大的数据集;然而，注释更大的数据集是具有挑战性的。在这里，我们提出一个 自动化方法预测ODX复发风险，无需注释。如果成功，这 方法将具有广泛的应用，包括但不限于 廉价的，基于网络的工具，预测ODX在发展中国家或农村地区与互联网 在标准Oncotype Dx检测成本过高或需要太长时间才能获得的情况下， 此外，我们的方法将用于临床研究，其中有价值的肿瘤组织可以 通过标准H& E染色切片获取相关研究数据，从而保存数据。

Efficient and cost-effective breast cancer risk stratification using whole-slide histopathology images Breast cancer prognosis depends highly on receptor status, as optimal treatment depends on the presence or absence of overexpression of estrogen, progesterone, or HER-2/neu receptors. To prevent over-treating patients with chemotherapy, it is crucial to quantify the risk of recurrence for estrogen receptor (ER) positive (ER+), HER2 negative (HER2-) breast cancer. A common assessment method to meet this need is the Oncotype DX (ODX) Recurrence Score. Unfortunately, ODX and similar gene assays are expensive, time-consuming, and tissue destructive. As an alternative, we propose estimating the ODX recurrence score using routine, ubiquitous, and inexpensive hematoxylin and eosin (H&E) staining of biopsies. There are other efforts to predict ODX recurrence risk from H&E. These automated methods detect histological primitives (e.g., nuclei) often in specific, also automatically detected, anatomical regions (e.g., ducts, tubules, lumen, epithelium, and stroma). Classification is performed into two or three risk categories, often collapsing two categories into one. The performance of these models is promising but still modest. One way to improve the performance of the models is to train on larger datasets; however, annotating larger datasets is challenging. Here, we propose an automated method to predict ODX recurrence risk without annotations. If successful, this method would have a wide range of applications, including but not limited to the availability of an inexpensive, web-based tool to predict ODX in developing countries or rural areas with internet access where standard Oncotype Dx assay would be cost-prohibitive or take too long to obtain. Furthermore, our method would find use in clinical research where valuable tumor tissue could be saved by obtaining correlative research data based on standard H&E-stained slides.