Predicting Response of Triple-Negative Breast Cancer to Neoadjuvant Chemotherapy Using a Deep Convolutional Neural Network-Based Artificial Intelligence Tool

Abstract

Purpose: Achieving a pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) is associated with improved patient outcomes in triple-negative breast cancer (TNBC). Currently, there are no validated predictive biomarkers for the response to NAC in TNBC. We developed and validated a deep convolutional neural network-based artificial intelligence (AI) model to predict the response of TNBC to NAC.

Materials and methods: Whole-slide images (WSIs) of hematoxylin and eosin-stained core biopsies from 165 (pCR in 60 and non-pCR in 105) and 78 (pCR in 31 and non-pCR in 47) patients with TNBC were used to train and validate the model. The model extracts morphometric features from WSIs in an unsupervised manner, thereby generating clusters of morphologically similar patterns. Downstream ranking of clusters provided regions of interest and morphometric scores; a low score close to zero and a high score close to one represented a high or low probability of response to NAC.

Results: The predictive ability of AI score for the entire cohort of 78 patients with TNBC ascertained by receiver operating characteristic analysis demonstrated an area under the curve (AUC) of 0.75. The AUC for stages I, II, and III disease were 0.88, 0.73, and 0.74, respectively. Using a cutoff value of 0.35, the positive predictive value of the AI score for pCR was 73.7%, and the negative predictive value was 76.2% for non-pCR patients.

Conclusion: To our knowledge, this study is the first to demonstrate the use of an AI tool on digitized hematoxylin and eosin-stained tissue images to predict the response to NAC in patients with TNBC with high accuracy. If validated in subsequent studies, these results may serve as an ancillary aid for individualized therapeutic decisions in patients with TNBC.

FIG 1.

Study workflow for determination of AI prediction score. AI, artificial intelligence; AUC, area under the curve; H&E, hematoxylin and eosin; NACT, neoadjuvant chemotherapy; ROI, regions of interest; TNBC, triple-negative breast cancer; WSI, whole-slide image.

FIG 2.

(A) ROC analysis demonstrating the performance of the deep CNN-based AI model to predict response of the entire testing cohort of 78 patients with triple-negative breast cancer using the definition of pathologic complete response as no residual invasive tumor with inclusion of residual in situ tumor after NAC showing the AUC of 0.75. (B) ROC analysis of the AI prediction model in patients with stage I disease shows an AUC of 0.88. (C) ROC analysis of patients with stage II disease shows an AUC of 0.73. (D) ROC analysis of patients with stage III disease shows an AUC of 0.74. AI, artificial intelligence; AUC, area under the curve; CNN, convolutional neural network; NAC, neoadjuvant chemotherapy; ROC, receiver operating characteristic.

FIG 3.

(A) ROC analysis demonstrating the performance of the deep CNN-based AI model to predict response of the entire testing cohort of 78 patients with triple-negative breast cancer using the definition of pathologic complete response as no residual in situ and invasive primary tumor and no evidence of metastatic tumor excluding isolated tumor cells in axillary lymph nodes after NAC showing the AUC of 0.73. (B) ROC analysis of the AI prediction model in patients with stage I disease shows an AUC of 0.88. (C) ROC analysis of patients with stage II disease shows an AUC of 0.70. (D) ROC analysis of patients with stage III disease shows an AUC of 0.74. AI, artificial intelligence; AUC, area under the curve; CNN, convolutional neural network; NAC, neoadjuvant chemotherapy; ROC, receiver operating characteristic.

FIG 4.

(A) H&E-stained core biopsy sample of a triple-negative breast cancer procured before NAC showing (B) high Nottingham histologic grade. (C) The tiling and labeling of a whole-slide digital image of the H&E-stained section of the core biopsy to include features of the tumor, stroma, and TILs to obtain the final score is shown in panels. (D) The deep CNN-based artificial intelligence prediction score generated as the collective scores of ROI including those with a prediction of good response (blue colored tiles) and bad response (green colored tiles) indicates tumor heterogeneity. The final AI prediction score of this case was 0.95, indicating a high likelihood of no response to NAC. The prediction of probability of response to NAC was accurate, as evidenced by absence of response to NAC with residual tumor cellularity of at least 95%. (E) and (F) The tumor bed with residual tumor. CNN, convolutional neural network; H&E, hematoxylin and eosin; NAC, neoadjuvant chemotherapy; ROI, regions of interest; TIL, tumor infiltrating lymphocyte.