BREAST CANCER CLASSIFICATION FROM HISTOPATHOLOGICAL IMAGES USING EXPLAINABLE ARTIFICIAL INTELLIGENCE: A PATIENT-WISE EVALUATION OF CNN AND VISION TRANSFORMER ARCHITECTURES
INTERNATIONAL EUROASIA Congress on Scientific Researches and Recent Trends XV, Makkah, Suudi Arabistan, 3 - 10 Ekim 2026, cilt.15, sa.1, ss.131-138, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 15
- Basıldığı Şehir: Makkah
- Basıldığı Ülke: Suudi Arabistan
- Sayfa Sayıları: ss.131-138
- Ankara Üniversitesi Adresli: Evet
Özet
Introduction and Purpose: Automated classification of breast cancer from histopathological images may support diagnostic assessment. This study compared CNN and Vision Transformer architectures under a patient-wise data split in terms of classification performance, computational cost, and explainability. It also investigated the relationship between performance, computational efficiency, and explanation faithfulness. Materials and Methods: The 7,909 benign and malignant images from 82 patients in the BreakHis dataset were divided into training, validation, and test sets without placing images from the same patient in different sets. ResNet50, DenseNet121, EfficientNet-B3, ViT-B/16, Swin-T, CrossViT-15, and MaxViT-T were trained, and the checkpoint with the highest validation F1 score was selected for each model. Performance was evaluated at image and patient levels using accuracy, sensitivity, specificity, F1, and ROC-AUC. Computational requirements were also examined. Grad-CAM, Integrated Gradients, and KernelSHAP were applied to EfficientNet-B3 and MaxViT-T selected according to validation results. Explanation faithfulness was measured using 100 balanced test images. Results: CrossViT-15 achieved the highest threshold-based performance, with 92.41% accuracy, 94.67% F1, and 98.00% sensitivity. MaxViT-T showed the strongest class discrimination with a ROC-AUC of 97.45%. DenseNet121 was the most compact model, with 6.96 million parameters. Integrated Gradients demonstrated the strongest explanation faithfulness for both models in the XAI evaluation. Discussion and Conclusion: No single architecture was superior across all criteria. When patient-wise evaluation, computational cost, and multiple XAI methods were considered together, CrossViT-15, MaxViT-T, and DenseNet121 emerged as preferable models for different priorities.