Breast mass classification in 3D ABUS based on Laplace-Beltrami spectra and dual path CNN

Abstract

Breast cancer is the most common cancer among women and remains a leading cause of cancer-related mortality worldwide. Accurately classifying breast masses as benign or malignant is crucial for guiding treatment and reducing unnecessary interventions. In this paper, we propose a hybrid deep learning-based classification framework for automated three-dimensional breast ultrasound (3D ABUS) images. The system integrates three classification paths: Support Vector Machine (SVM), Extremely Randomized Trees (Extra Trees), and a novel deep neural network. The SVM and Extra Trees classifiers utilize handcrafted features, including radiomic descriptors and Laplace-Beltrami eigenvalues. In these models, to reduce dimensionality of the Laplace-Beltrami features and prevent overfitting, Isomap is employed for nonlinear dimensionality reduction. The proposed neural network processes a 3D patch around the mass and the corresponding mask through two parallel paths, utilizing convolutional and max-pooling layers. The extracted features from both branches are concatenated with the complete Laplace-Beltrami feature vector before being classified by fully connected layers. To combine the outputs of all three base models, we employ a histogram-based gradient-boosting stacking classifier. This meta-classifier learns nonlinear dependencies between classifiers and enhances the overall performance. Experimental evaluation was conducted on the public TDSC-ABUS dataset, comprising 200 annotated breast volumes. The training/validation set includes 75 malignant and 55 benign cases, while the test set contains 40 malignant and 30 benign cases. On the test set, the proposed system achieves 84.29% accuracy, 93.50% AUC, 97.50% sensitivity, and 87.64% F1-score. Compared to the best competing method, it improves accuracy by 8.58% and AUC by 4.58%.