The clinical validity of radiomics-based prediction of molecular subtypes in breast cancer from digital mammary tomosynthesis

TL;DR

This study uses digital breast tomography and machine learning to predict breast cancer molecular subtypes, creating a diagnostic model with clinical potential.

Contribution

A novel radiomics-based nomogram model for predicting breast cancer molecular subtypes using digital breast tomography and machine learning.

Findings

Random Forest was the best classifier with the highest AUC, accuracy, and F1 score.

A nomogram model was developed showing increased odds of breast cancer with higher total scores.

Key features of breast cancer were identified using image omics and machine learning.

Abstract

To explore the use of digital breast tomography (DBT) imaging omics in developing breast cancer (BC) diagnostic models to identify molecular subtype characteristics of BC.

A retrospective analysis was conducted on 433 DBT images. Candidate features were extracted, and least absolute shrinkage and selection operator (LASSO) regression model was established. Within the training set, machine learning (ML) models were constructed, and their predictive performance was evaluated using receiver operating characteristic (ROC) curves and confusion matrixes in the test set, thereby screening the best predictive classifier. Univariate and multivariate Cox regression analyses were conducted to obtain key characteristics of nomogram modeling, correction and decision curve analysis (DCA) were used to evaluate the clinical potential of this model.

The LASSO selected 14 features. Random Forest (RF) had the highest AUC value, the highest accuracy, sensitivity, recall rate and F1 score on the training set and test set, and was the best classifier. A nomogram model was established. The odds ratio (OR) of BC patients increased with the increase of the total score.

The key features of BC were revealed by image omics and ML models, and a nomogram model with diagnostic value was constructed.

Figures