Contrast-enhanced mammography-based interpretable machine learning model for the prediction of the molecular subtype breast cancers

Abstract

Objective: This study aims to establish a machine learning prediction model to explore the correlation between contrast-enhanced mammography (CEM) imaging features and molecular subtypes of mass-type breast cancer.

Materials and methods: This retrospective study included women with breast cancer who underwent CEM preoperatively between 2018 and 2021. We included 241 patients, which were randomly assigned to either a training or a test set in a 7:3 ratio. Twenty-one features were visually described, including four clinical features and seventeen radiological features, these radiological features which extracted from the CEM. Three binary classifications of subtypes were performed: Luminal vs. non-Luminal, HER2-enriched vs. non-HER2-enriched, and triple-negative (TNBC) vs. non-triple-negative. A multinomial naive Bayes (MNB) machine learning scheme was employed for the classification, and the least absolute shrink age and selection operator method were used to select the most predictive features for the classifiers. The classification performance was evaluated using the area under the receiver operating characteristic curve. We also utilized SHapley Additive exPlanation (SHAP) values to explain the prediction model.

Results: The model that used a combination of low energy (LE) and dual-energy subtraction (DES) achieved the best performance compared to using either of the two images alone, yielding an area under the receiver operating characteristic curve of 0.798 for Luminal vs. non-Luminal subtypes, 0.695 for TNBC vs. non-TNBC, and 0.773 for HER2-enriched vs. non-HER2-enriched. The SHAP algorithm shows that "LE_mass_margin_spiculated," "DES_mass_enhanced_margin_spiculated," and "DES_mass_internal_enhancement_homogeneous" have the most significant impact on the model's performance in predicting Luminal and non-Luminal breast cancer. "mass_calcification_relationship_no," "calcification_ type_no," and "LE_mass_margin_spiculated" have a considerable impact on the model's performance in predicting HER2 and non-HER2 breast cancer.

Conclusions: The radiological characteristics of breast tumors extracted from CEM were found to be associated with breast cancer subtypes in our study. Future research is needed to validate these findings.

Keywords: Breast cancer; Contrast-enhanced mammography; Interpretable machine learning model; Molecular subtype; Multinomial naive Bayes; SHapley additive exPlanation.

Fig. 1

Flowchart of the study population

Fig. 2

A woman with palpable left breast mass. Left craniocaudal (a) and mediolateral oblique view (c) of LE showed an irregular hyperdense mass in the upper inner quadrant. DES images (b, d) showed an irregular mass with homogeneous internal enhancement. She underwent breast surgery, and the final pathological diagnosis was invasive ductal breast carcinomas of triple negative

Fig. 3

A woman with palpable right breast mass. Right craniocaudal (a) and mediolateral oblique view (c) of LE showed an irregular hyperdense mass in the upper outer quadrant, with fine pleomorphic calcifications in and anteriorly of the mass. DES images (b, d) showed an irregular mass with heterogeneous internal enhancement, irregular margin, and marked non-mass enhancement of calcification in front of the mass. She underwent breast surgery, and the final pathological diagnosis was HER2-enriched breast cancer

Fig. 4

ROC curves from the MNB model for predicting the molecular subtype of breast cancer patients. Curves were constructed from radiological features from LE alone, DES alone and a combination of LE and DES views in the testing set, respectively. Notes: AUC, area under the curve; ROC, receiver operating characteristic; LE, low-energy; DES, dual-energy subtraction. a) Luminal vs. non-Luminal, b) HER2-enriched vs. non-HER2-enriched, and c) TNBC vs. non-TNBC

Fig. 5

Predictor importance as considered by a machine learning trained and validated on 241 patients. The relative contributions of variables for progression prediction quantified with the mean of the absolute Shapley additive explanation (SHAP) values. a) Luminal vs. non-Luminal, b) HER2-enriched vs. non-HER2-enriched, and c) TNBC vs. non-TNBC

Fig. 6

A woman with palpable left breast mass. Left oblique magnification (a) and local amplified photographing (c) of the upper showed an irregular hyperdense mass, with on spiculated margin and no definite calcification inside. DES images (b, d) also showed an irregular mass with spiculated margin, and homogeneous internal enhancement. The final pathological diagnosis was invasive ductal breast carcinomas of luminal subtype. In addition, SHAP analysis was performed on the patient’s imaging findings (e). Red arrows indicate positive Shapley values increasing the probability and blue arrows indicate negative values decreasing the probability