Abstract
PURPOSE: The aim of this study is to investigate a new method that combines radiological and pathological breast cancer information to predict discrepancies in pathological responses for individualized treatment planning. We used baseline multiparametric magnetic resonance imaging and hematoxylin and eosin-stained biopsy slides to extract quantitative feature information and predict the pathological response to neoadjuvant chemotherapy in breast cancer patients. METHODS: We retrospectively collected data from breast cancer patients who received neoadjuvant chemotherapy in our hospital from August 2016 to January 2018; multiparametric magnetic resonance imaging (contrast-enhanced T1-weighted imaging and diffusion-weighted imaging) and whole slide image of hematoxylin and eosin-stained biopsy sections were collected. Quantitative imaging features were extracted from the multiparametric magnetic resonance imaging and the whole slide image were used to construct a radiopathomics signature model powered by machine learning methods. Models based on multiparametric magnetic resonance imaging or whole slide image alone were also constructed for comparison and referred to as the radiomics signature and pathomics signature models, respectively. Four modeling methods were used to establish prediction models. Model performances were evaluated using receiver operating characteristic curve analysis and the area under the curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. RESULTS: The radiopathomics signature model had favourable performance for the prediction of pathological complete response in the training set (the best value: area under the curve 0.83, accuracy 0.84, and sensitivity 0.87), and in the test set (the best value: area under the curve 0.91, accuracy 0.90, and sensitivity 0.88). In the test set, the radiopathomics signature model also significantly outperformed the radiomics signature (the best value: area under the curve 0.83, accuracy 0.64, and sensitivity 0.62), pathomics signature (the best value: area under the curve 0.60, accuracy 0.74, and sensitivity 0.62) (p > 0.05). Decision curve analysis and calibration curves confirmed the excellent performance of these prediction models in discrimination, calibration, and clinical usefulness. CONCLUSIONS: The results of this study suggest that radiopathomics, the combination of both radiological information regarding the whole tumor and pathological information at the cellular level, could potentially predict discrepancies in pathological response and provide evidence for rational treatment plans.