A Deep Learning Approach for Fusing Contrast-Enhanced Mammogram Features for Breast Tumor Detection an‎d Classification

10/23/2025 12:00:00 AM

Breast cancer is one of the most common cancers among women, an‎d its early diagnosis plays a vital role in improving treatment outcomes. Contrast-Enhanced Mammography (CEM) is a novel imaging modality that, by simultaneously providing two types of images—low-energy an‎d subtracted—holds high potential for the accurate detection of masses. However, many existing methods do not fully leverage the synergistic potential of these two image types. This thesis presents a comprehensive deep learning-based approach for fusing features from CEM images for the detection an‎d classification of breast masses. The core innovation of this research is the design of dual-stream architectures: a "Dual-Stream RetinaNet" model for the detection task an‎d a "Dual-Stream EfficientNetV2-S" model for the classification task. These architectures process the low-energy an‎d subtracted images in parallel, fusing the extracted features to create a rich an‎d comprehensive representation of the lesion an‎d the image. Furthermore, for the first time, a dedicated preprocessing pipeline for this data type has been implemented, which includes a spatial registration step to ensure perfect alignment between the images an‎d a symmetric data augmentation strategy to maintain this alignment throughout the training process. The proposed models were eva‎luated on the public CDD-CESM dataset. The detection model achieved an Average Precision (AP50) of 65.32% under a 5-fold cross-validation protocol. The classification model achieved an F1-score of 91.43%, an accuracy of 91.18%, an‎d an Area Under the Curve (AUC) of 0.9729. The results demonstrate that the feature fusion approach, via the dual-stream architecture, effectively combines complementary information, leading to a significant improvement in diagnostic accuracy. This research demonstrates the high potential of this method as a powerful assistive tool for radiologists in the analysis of CEM images an‎d the early detection of breast cancer.

سرطان پستان , ماموگرافي با تباين بهبود يافته , رتينانت جريان دوگانه , آشكارسازي و رده‌بندي ضايعات , ادغام ويژگي

Breast Cancer , Contrast-Enhanced Mammography (CEM) , Dual-Stream RetinaNet , Lesion Detection an‎d Classification , Feature Fusion

Amirhossein Haji Jafari

Mohsen Soryani