Quantitative digital histopathology and machine learning to predict pathological complete response to chemotherapy in breast cancer patients using pre-treatment tumor biopsies

Abstract

Complete pathological response (pCR) to neoadjuvant chemotherapy (NAC) is a prognostic factor for breast cancer (BC) patients and is correlated with improved survival. However, pCR rates are variable to standard NAC, depending on BC subtype. This study investigates quantitative digital histopathology coupled with machine learning (ML) to predict NAC response a priori. Clinicopathologic data and digitized slides of BC core needle biopsies were collected from 149 patients treated with NAC. The nuclei within the tumor regions were segmented on the histology images of biopsy samples using a weighted U-Net model. Five pathomic feature subsets were extracted from segmented digitized samples, including the morphological, intensity-based, texture, graph-based and wavelet features. Seven ML experiments were conducted with different feature sets to develop a prediction model of therapy response using a gradient boosting machine with decision trees. The models were trained and optimized using a five-fold cross validation on the training data and evaluated using an unseen independent test set. The prediction model developed with the best clinical features (tumor size, tumor grade, age, and ER, PR, HER2 status) demonstrated an area under the ROC curve (AUC) of 0.73. Various pathomic feature subsets resulted in models with AUCs in the range of 0.67 and 0.87, with the best results associated with the graph-based and wavelet features. The selected features among all subsets of the pathomic and clinicopathologic features included four wavelet and three graph-based features and no clinical features. The predictive model developed with these features outperformed the other models, with an AUC of 0.90, a sensitivity of 85% and a specificity of 82% on the independent test set. The results demonstrated the potential of quantitative digital histopathology features integrated with ML methods in predicting BC response to NAC. This study is a step forward towards precision oncology for BC patients to potentially guide future therapies.

Figure 1

(a) An example of tumor bed annotation on a representative segment of the core on WSI (2188 × 4124 pixels), (b) the non-overlapping 768 × 768 pixel tiles extracted from the tumor region with the excluded tiles shaded (less than 50% tumor or more than 10% white background), (c) an extracted tile with 100% tumor tissue, and (d) the generated binary mask of the nuclei in the tile.

Figure 2

The importance gain score of the first 15 features with highest contribution to the predictive model for different feature subsets: (a) clinical, (b) morphological, (c) intensity-based, (d) texture, (e) graph-based, (f) wavelet, and (g) all features. The green bars are associated with the features included in the NAC response biomarker in each experiment.

Figure 3

Box plots of the selected features for the pCR and non-pCR cohorts of the training set obtained in the seven experiments conducted using different feature subsets: (a) clinical, (b) morphological, (c) intensity-based, (d) texture, (e) graph-based, (f) wavelet, and (g) all features. The feature values are normalized in the range of 0 and 1. The order of features in each plot is the same as that of the associated plot in Fig. 2.

Figure 4

Receiver operating characteristic (ROC) curves on the independent test set for the predictive models developed with the selected features obtained in different experiments. In the last experiment and from all feature subsets, 7 wavelet and graph-based features were selected.