Developing Digital Pathology Biomarkers for Response to Neoadjuvant and Adjuvant Chemotherapy in Breast Cancer

点击翻译按钮获取中文摘要

Project Summary Although tremendous strides have been made in uncovering the biology of breast cancer, selection of chemotherapy regimens for early breast cancer is based predominantly on receptor status and stage. However, numerous other factors are associated with response, including gene expression patterns and tumor genetics, but these are not uniformly available for patients. Hematoxylin and eosin stained pathology is routinely obtained for all patients with breast cancer, and contains a wealth of information beyond grade. For example, the pattern and amount of tumor infiltrating lymphocytes has long been recognized as a predictor of response to chemotherapy, but quantification is challenging. Deep learning is an emerging discipline with particular promise in the domain of image recognition, wherein models can learn from repeated exposure to sample images to recognize any candidate features of interest. Using deep learning, our group and others have successfully used histology to predict a variety of tumor specific factors linked with response to treatment, including receptor status, gene expression patterns, driver mutations, and tumor infiltrating lymphocytes. These features can be accurately detected at point of care, without the extended turn-around time and expense associated with specialized molecular testing. We hypothesize that deep learning on histology can identify novel morphologic and spatial features of breast cancer tumors that in turn can predict response to chemotherapy in early breast cancer. We will take advantage of a rich institutional cohort of over 600 patients who received neoadjuvant chemotherapy and over 2000 patients with long term survival data to curate a well annotated database suited for deep learning on digital histology. Our patient cohort also features diverse demographics with inclusion of minority patients often underrepresented in public datasets, ensuring applicability of our findings to all patients with breast cancer. We will use this dataset to develop a deep learning histologic biomarker of chemotherapy response in early stage breast cancer. This deep learning biomarker will be compared to standard markers of response to determine if deep learning on histology provides independent predictive value, allowing better identification of candidates for intensification or de-intensification of standard anthracycline and taxane based chemotherapy.