A comparative decision support framework for breast ultrasound classification using transfer learning, hybrid deep representations, and BPSO-based feature selection
EXPERT SYSTEMS WITH APPLICATIONS, cilt.332, 2027 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 332
- Basım Tarihi: 2027
- Doi Numarası: 10.1016/j.eswa.2026.133506
- Dergi Adı: EXPERT SYSTEMS WITH APPLICATIONS
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, Public Affairs Index, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Kayseri Üniversitesi Adresli: Evet
Özet
Breast cancer is one of the most common malignancies among women, and early diagnosis is critical for improving survival. Breast ultrasonography is widely used because it is non-invasive, cost-effective, and suitable for dense breast tissue. However, low contrast, speckle noise, and operator dependence complicate interpretation and increase the need for reliable computer-aided decision-support systems. This study proposes a comparative decision-support framework for three-class breast ultrasound classification (benign, malignant, and normal) on the BUSI dataset. The framework evaluates a baseline CNN, ResNet50, DenseNet121, Effi-cientNetV2B0, MobileNetV2, and a hybrid CNN-Transformer under the same data split, input size, training setting, and evaluation protocol. Moreover, 1024-dimensional deep features are selected using Binary Particle Swarm Optimization (BPSO) and classified with a multiclass support vector machine (SVM). To prevent data leakage, BPSO is performed only on the training set using 5-fold cross-validation, while the final SVM is trained on the combined training and validation sets. Class imbalance is addressed with inverse-frequency weighting. Statistical reporting includes 95% bootstrap confidence intervals, Matthews correlation coefficient, Cohen's kappa, geometric mean, sensitivity analyses for BPSO/SVM hyperparameters, and computational cost measures is reported for every backbone. Experimental results show that the hybrid CNN-Transformer achieves the highest end-to-end accuracy (0.7917) and ROC-AUC Macro-OVR (0.9203), whereas ResNet50 obtains the best end-to-end Macro-F1 (0.7780), balanced accuracy (0.8107), and malignant sensitivity (0.8438). Among hybrid pipelines, ResNet50+BPSO+SVM and EfficientNetV2B0+BPSO+SVM achieve the highest accuracy (0.8667), while ResNet50+BPSO+SVM yields the best Macro-F1 (0.8451). Overall, deep feature-based BPSO+SVM pipelines improve feature efficiency and decision performance for breast ultrasound classification.