GTO-DEEPSCN: OPTIMIZATION OF DEEP STOCHASTIC CONFIGURATION NETWORKS USING THE ARTIFICIAL GORILLA TROOPS OPTIMIZER FOR BREAST MRI CLASSIFICATION
16th International Azerbaijan Congress on Life, Engineering, Mathematical, and Applied Sciences, Baku, Azerbaycan, 20 - 22 Eylül 2026, ss.166-173, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.30546/19023.978-9952-610-91-5.2026.100.1148
- Basıldığı Şehir: Baku
- Basıldığı Ülke: Azerbaycan
- Sayfa Sayıları: ss.166-173
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Breast cancer is one of the most commonly diagnosed cancers among women, and early and accurate diagnosis is essential for improving patient prognosis. Although artificial intelligence-based methods have achieved promising results in medical image classification, many deep learning models suffer from high computational cost, numerous trainable parameters, and extensive hyperparameter tuning requirements. Stochastic Configuration Networks (SCNs) offer a relatively lower-complexity classification framework; however, their random parameter generation may cause performance variability across runs. In this study, a hybrid GTO-DeepSCN model integrating a Deep Stochastic Configuration Network with the Artificial Gorilla Troops Optimizer (GTO) is proposed to improve parameter determination and classification performance. Discriminative deep features were extracted using a fine-tuned EfficientNetB4 model and reduced through a hybrid feature selection strategy before being provided to the GTO-DeepSCN classifier. The GTO algorithm optimizes the input weights and bias values of DeepSCN nodes, reducing the influence of random parameter generation. Hidden layers, normalization, and skip connections were also incorporated to improve learning stability and information flow. Experiments were conducted on a breast MRI dataset containing approximately 25,000 benign and malignant images. The proposed model was comparatively evaluated against the standard DeepSCN. GTO-DeepSCN achieved a test accuracy of 93.29%, ROC-AUC of 0.9813, sensitivity of 97.11%, NPV of 96.87%, and macro-average F1-score of 93.28%. Compared with the standard DeepSCN, sensitivity increased from 90.57% to 97.11%, while false-negative malignant samples decreased from 170 to 52. Overall, the findings indicate that GTO-based optimization improves DeepSCN classification performance and provides a promising approach for benign and malignant breast MR image classification.