Neural Network-Based Analysis of Mammography Images for Identifying Breast Cancer Histological Subtypes

Breast cancer (BC) is a complex disease with multiple histological subtypes that exhibits distinct biological and clinical characteristics. Accurate identification of these subtypes is crucial for the implementation of personalized treatment strategies and the improvement of patient outcomes. This study aimed to leverage the potential of neural networks and mammography data to identify specific histological subtypes of BC, with a focus on distinguishing between in situ and invasive carcinoma. We used the Digital Database for Screening Mammography (DDSM) and its curated subset, CBIS-DDSM, which provides mammography images of normal and malignant cases with verified pathological information. A convolutional neural network (CNN) architecture was designed to extract relevant features from mammography images. The model was trained using data augmentation techniques to enhance diversity and mitigate overfitting, resulting in training and validation accuracies ranging from 0.98 to 1.0. The features extracted by the trained model were then used for clustering analysis using Kmeans on PCA (KM-PCA), which identified nine well-separated clusters with minimal overlap. The clustering results were validated using a subset of mammography images with confirmed invasive carcinoma types from The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA) dataset. Two validation approaches were incorporated to strengthen the hypothesis and further validate the results. Overall, this study demonstrates the potential of combining neural networks and mammography data to accurately identify BC histological subtypes, which could pave the way for personalized and effective treatment strategies.