Accurate phenotyping of luminal A breast cancer in magnetic resonance imaging: A new 3D CNN approach

Abstract

Breast cancer (BC) remains a predominant and deadly cancer in women worldwide. By 2040, projections indicate that more than 3 million new cases of breast cancer will emerge annually, culminating in more than 1 million deaths worldwide. Early detection and accurate diagnosis of BC are critical factors that influence treatment success and patient outcomes. During the past three decades, several medical imaging modalities, such as X-ray Mammography (MG), Ultrasound (US), Computer Tomography (CT), Magnetic Resonance Imaging (MRI), and Digital Tomosynthesis (DT) have been explored to support radiologists/physicians in clinical decision-making workflows for the detection, diagnosis, treatment, and monitoring of BC. Magnetic resonance imaging (MRI) is an advanced imaging modality that provides detailed information on the structure and function of breast tissue. In particular, MRI may be crucial to discern the phenotype of BC, as each subtype has a different prognosis and requires different treatment strategies. This study aims to explore deep learning models for classifying/diagnosing BC phenotypes. As a main contribution, we propose a new 3D convolutional neural network (CNN) model based on quantitative medical imaging biomarkers (QIB) obtained from MRI data to diagnose the luminal A subtype (LA) of BC. LA is a subtype characterized by positive hormone receptor expression and negative HER2 expression. It uses a binary classification strategy to distinguish between pathological luminal A and non-luminal A lesions by analyzing 3D volumetric MRI images. The proposed method allows the extraction and analysis of spatial information, which is essential to accurately diagnose BC, especially for the LA subtype, taking into account their specific morphological characteristics. Our goal is to improve accuracy and efficacy in the diagnosis of the LA phenotype of BC and to contribute to the development of personalized treatment plans for patients. To develop and evaluate the performance of the proposed method, we used a benchmarking public domain MRI-based BC dataset (Duke-Breast-Cancer-MRI). To address the imbalance in the data set, we implemented a class weighting strategy during model training. In experimental settings, we achieved an AUC score of 0.9614 and a $F_1$ score of 0.9328, outperforming state-of-the-art methods, including ResNet-152. These results demonstrate the potential of our work to significantly improve the diagnosis of the luminal A phenotype of breast cancer, paving the way for more accurate and personalized treatment strategies.