Machine learning-based classification model to differentiate subtypes of invasive breast cancer using MRI

TL;DR

This paper presents a machine learning model that uses MRI scans to distinguish between two types of invasive breast cancer based on the contralateral breast's shape and size.

Contribution

A novel machine learning model using contralateral breast morphology to classify IDC and ILC with 79% accuracy.

Findings

The model achieved 79% accuracy and an AUC of 0.851 in differentiating IDC and ILC.

Contralateral breast volume, surface area, and density were key features for classification.

Morphological features of the contralateral breast are important for subtype differentiation.

Abstract

Breast cancer is considered one of the most lethal diseases among women worldwide. Invasive Ductal Carcinoma (IDC) and Invasive Lobular Carcinoma (ILC) are the two most prominent subtypes of breast cancer. They differ in epidemiology, molecular alterations, and clinicopathological features. Patient treatment and management also differ due to these variations.

The study aimed to develop a predictive model to differentiate IDC and ILC using machine learning techniques based on the morphological features of the contralateral breast. Methods- 143 magnetic resonance imaging (MRI) images were sourced from the “DUKE Breast-Cancer” collection on the Cancer Imaging Archive website. Regions of interest were drawn on each slice to compute the morphological features of the contralateral breast using the 3D Slicer application. Supervised learning methods were applied to the morphological features to build a predictive model incorporating a Random Forest Classifier to differentiate IDC and ILC. Hyperparameters were tuned to optimize the model.

The model was able to differentiate IDC and ILC with an accuracy of 79% and an Area Under the Curve of 0.851 on the Receiver Operating Characteristic Curve. Among the morphological features, the total volume of the contralateral breast, surface area of the contralateral breast, breast density, and the ratio of the total volume of the contralateral breast to its surface area had higher F-scores, indicating that the dimensions of the contralateral breast could be an important factor in differentiating IDC and ILC.

This study successfully developed and optimized a predictive model based on breast morphological features to differentiate IDC and ILC using machine learning methods.

Figures