# Predicting histological grade in invasive ductal carcinoma of the breast: a radiomics-based machine learning model using DCE-MRI

## TL;DR

This study uses MRI images and machine learning to predict the severity of breast cancer before surgery, helping doctors make better treatment decisions.

## Contribution

A radiomics-based logistic regression model is proposed for non-invasive preoperative prediction of histological grade in breast cancer.

## Key findings

- 22 key radiomics features were selected from DCE-MRI images for model training.
- Logistic regression outperformed other models with an AUC of 0.795 in predicting histological grade.
- The model shows potential for supporting individualized clinical decision-making in breast cancer.

## Abstract

To investigate the feasibility analysis of predicting the pathological differentiation grade of breast invasive ductal carcinoma based on DCE-MRI imaging histology.

198 patients with breast invasive ductal carcinoma who underwent preoperative enhanced MRI were retrospectively collected from January 2019 to October 2024.According to Nottingham histologic grading, 108 cases were divided into a high-grade group and 90 cases into an intermediate-low-grade group, which were randomly divided into 148 cases of the training group and 50 cases of the validation group according to a 3:1 ratio. The 3D slicer software was applied to extract the image histological features of the region of interest, and five models, namely, decision tree, Gaussian plain Bayes, logistic regression, random forest, and AdaBoost, were constructed by filtering the features with intragroup correlation coefficients and the minimum absolute contraction and selection operators. Compare the area under the work characteristic curve of subjects in the validation group and select the best model. The performance of the best model validation group was evaluated, the clinical usability of the best model was examined using decision curves, and the accuracy of the predictive model was visualized using calibration curves.

After rigorous stability and redundancy screening, 22 key radiomics features were selected from DCE-MRI images. Multiple machine learning models trained based on these features were evaluated for their predictive performance on the validation set. The logistic regression model achieved the highest AUC value of 0.795 (95% confidence interval: 0.664-0.927), outperforming other models such as random forest (AUC = 0.700), Gaussian naive Bayes (AUC = 0.700), AdaBoost (AUC = 0.718), and decision tree (AUC = 0.587). Consequently, the logistic regression model was ultimately selected as the optimal model.

The DCE-MRI radiomics model based on Logistic Regression can non-invasively and effectively predict the histological grade of IDC preoperatively, offering valuable potential for supporting individualized clinical decision-making.