- Home
- Archives
- 2026
- Vol. 6 Issue 6
- Article
Hybrid Deep Feature Learning for Benign–Malignant Breast Lesion Classification in Ultrasound Images
Authors
1 Department of Artificial Intelligence, College of Computer Science and Information Technology, Kirkuk University, Kirkuk, Iraq.
Article Information
- Published In
- Volume 6, Issue 6 (2026)
- Published Online
- November-December 2026
- Pages
- 66-76
- Article Type
- Original Article
Abstract
View PDFBreast cancer is the most frequently diagnosed cancer in women, and distinguishing benign from malignant lesions on ultrasound (US) images remains a largely operator-dependent and subjective task. In this study, five deep learning models representing three architectural generations, namely classical convolutional neural networks (CNNs) (VGG16, ResNet50V2 and DenseNet121), a modern CNN (ConvNeXt-Tiny) and a hierarchical vision Transformer (Swin-Tiny), were compared on the publicly available BUS-BRA dataset within a two-stage framework comprising end-to-end fine-tuning and deep feature-based hybrid classification. To prevent data leakage, the 1875 images obtained from 1064 patients were divided into training, validation and test sets at the patient level. In the hybrid stage, 256-dimensional deep features extracted from each fine-tuned model were combined with five classical classifiers, namely support vector machine, random forest, k-nearest neighbours, logistic regression and gradient boosting, and both stages were evaluated on the same held-out test set of 375 images. In the end-to-end setting, ConvNeXt-Tiny achieved the highest performance, with an accuracy of 91.47% and an area under the receiver operating characteristic curve (AUC) of 0.953. Random forest-based hybrid models outperformed their end-to-end counterparts for all five backbones, improving accuracy by 3.46 to 5.06 percentage points, and the best overall result was obtained with the ConvNeXt-Tiny and random forest combination (accuracy 94.93%, sensitivity 92.56%, specificity 96.06% and AUC 0.982). In contrast, the k-nearest neighbours classifier reduced performance for every backbone. These findings indicate that the benefit of hybrid modelling depends strongly on the choice of classifier, and they underline the importance of a patient-level, leakage-free evaluation protocol for obtaining reliable performance estimates in breast US image analysis.
Related Articles
Share Article
*Instagram doesn't support direct link sharing from web. Copy the link and share it in your Instagram story or post.