Deep Learning Approaches for Tumor-Stroma Classification in Colorectal Cancer Pathology
8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Türkiye, 21 - 23 Mayıs 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ichora69329.2026.11537022
- Basıldığı Şehir: Ankara
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: Colorectal cancer, Deep learning, Ensemble learning, Histopathological image analysis, Tumor-stroma ratio
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
In this study, five deep learning models - ResNet50, a fine-tuned version of ResNet50, MobileNetV2, MobileNetV3, and a custom-built CNN - were evaluated to assist tumorstroma ratio (TSR) assessment in histopathological images of colorectal cancer. All models were evaluated on a three-class (tumor, stroma, other) image patch-level classification problem. Performance evaluation was conducted using accuracy along with class-based precision, recall, and F1-score metrics. In addition, an ensemble strategy combining the softmax outputs of individual models was evaluated. The experimental results demonstrate that the fine-tuned ResNet50 model achieves the highest accuracy among the individual architectures. However, the proposed ensemble approach achieved the highest overall performance, with 96% accuracy and an F1-score of 0.99 for the tumor class. While MobileNet-based architectures provide well-balanced performance with reduced computational costs, the Custom CNN architecture attained high sensitivity for the tumor class, primarily due to the weighting strategy used to address class imbalance.