Advanced Deep Learning Model for Microcalcification Detection in Digital Breast Tomosynthesis

Abstract

Microcalcification (MC) detection is a critical aspect of early breast cancer diagnosis, as these tiny calcium deposits can indicate malignant changes in breast tissue. Conventional techniques, like handcrafted feature extraction and machine learning (ML) models, face challenges to detect MC due to their microscopic size, subtle appearance, and overlapping tissue structures in mammographic images. Although deep learning (DL) approaches like convolutional neural networks (CNNs) have enhanced detection, they often fail to capture sequential dependencies and spatial hierarchies, reducing their effectiveness. This study proposes a hybrid DL model integrating Long Short-Term Memory (LSTM) and Capsule Networks (CapsNet) to enhance MC detection in digital breast tomosynthesis (DBT) images. LSTM captures sequential dependencies within structured feature representations, while CapsNet preserves spatial hierarchies and orientation relationships, leading to improved classification accuracy. The proposed framework was trained and evaluated on a mammography dataset from Kaggle, achieving 96.45% accuracy, 97.85% precision, 95.12% recall, and a 96.47% F1-score, signifying its efficiency in distinguishing MCs from non-MCs. These findings emphasize the robustness and reliability of the LSTM-CapsNet model, making it a reliable tool for improving the accuracy of automated MC detection and helping in early breast cancer diagnosis.