Abstract P5-02-06: Improved histologic grading of breast cancer by a novel deep learning-based model

Abstract

Abstract Breast cancer histologic grade is a well-established prognostic factor utilized in clinical decision making. In current clinical practice grading is conducted manually by pathologists and this procedure is associated with a substantial inter-observer variability. Furthermore, patients classified as histologic grade 2 has been reported to exhibit an intermediate recurrence risk, resulting in less prognostic value. With the aim of improving patient stratification, we have developed a model-based approach for histological grading using deep convolutional neural networks (CNNs). In this study, we developed a CNN model for improved histologic grading, with a focus on further stratification of grade 2 patients into high and low risk groups. Histopathology images from the Clinseq breast cancer study and the Cancer Genome Atlas (TCGA) containing 730 patients were scanned at 40X magnification and tiled into patches at 20X. In total, we obtained 5.3 million patches of size 299 × 299 pixels. Using image annotations of invasive cancer regions in Clinseq study, we trained a deep CNN model to segment cancer regions in TCGA data. Subsequently, the cancer regions from both Clinseq and TCGA were used to optimize a second deep CNN model for classification of grade 1 and 3, this model was applied to re-classify 291 patients with histologic grade 2 into low and high risk. The CNN model classified cancer and non-cancer with area under the receiver operating characteristic curve (AUC) of 0.941. The second CNN model achieved an AUC of 0.910 in terms of classification of grade 1 and 3 tumors (cross-validation). For grade 2 tumors (independent test data), 184 were re-classified into the lower risk group whereas 107 were classified as high risk. The risk of recurrence for the grade 2 higher risk group is significantly higher than that in lower risk group, with an estimated hazard ratio of 2.86 (95% confidence interval: 1.21-6.69) after adjusting for age, tumor size, estrogen receptor status and lymph node status. In conclusion, we found that the deep CNN model demonstrated a high capability to distinguish between breast cancer with histologic grade 1 and 3. We also found that re-stratification of Grade 2 patients into high and low risk groups was significantly associated with risk of recurrence. Improved histological grading, and further risk stratification of grade 2 patients, by deep CNN models could contribute towards a reduction of both over- and under-treatment of breast cancer patients. Citation Format: Mattias Rantalainen, Yinxi Wang, Balázs Ácz, Stephanie Robertson, Johan Hartman. Improved histologic grading of breast cancer by a novel deep learning-based model [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P5-02-06.