A Comprehensive Framework for Multi-Class Breast Lesion Classification Using a Gaussian Process-Optimized Vgg16-Capsule Hybrid Model on Ultrasound Radio Frequency B-Mode Images Publisher



M Arab MAHSA ; A Fallah ALI ; S Rashidi SAEID ; M Mehdizadeh Dastjerdi MARYAM ; N Ahmadinejad NASRIN
Source: AUT Journal of Electrical Engineering Published:2025

Abstract

Effective breast cancer screening is essential for early detection and treatment. Ultrasound (US) radio frequency (RF) data offers a novel, equipment-independent approach. However, class imbalance and limited interpretability hinder its application in clinical practice. This study proposes a hybrid deep learning model combining a pre-trained convolutional neural network (CNN) based on VGG16 and capsule neural networks (CapsNets) to classify breast lesions. The model was evaluated using an RFTSBU dataset, comprising 220 data points from 118 patients, acquired on the SuperSonic Imagine Aixplorer® system with a linear transducer. To address data imbalance, the synthetic minority over-sampling technique (SMOTE) was employed to generate synthetic samples while preserving data distribution. Furthermore, Gaussian process (GP) was applied to fine-tune CapsNet hyperparameters, improving classification performance. Three experiments were conducted to classify breast lesions into two, three, and four classes: (I) CapsNet with balanced datasets based on class weight, (II) CapsNet with balanced datasets using SMOTE, and (III) CapsNet with hyperparameters optimized using GP on SMOTE-balanced datasets. The proposed model achieved average accuracies of 98.81%, 97.89%, and 95.94% for two-, three-, and four-class classifications, respectively. The hybrid VGG16-CapsNet model effectively addresses class imbalance and captures critical lesion attributes such as size, perspective, and orientation. Integrating GP optimization achieves superior accuracy in multi-class breast lesion classification. The proposed approach can serve as a valuable aid in breast tumor classification using US RF B-mode images. Its enhanced interpretability and efficiency enable clinicians to move beyond binary classification, facilitating the identification and differentiating a broader spectrum of breast lesions. © 2025 Elsevier B.V., All rights reserved.