Introducing Ranking Stabilization Epoch for Early Model Selection in Kidney Stone Detection from CT Images
2nd International Symposium on AI-Driven Engineering Systems, ISADES 2026, Hybrid, Mbale, Uganda, 19 - 20 Haziran 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/isades69945.2026.11608072
- Basıldığı Şehir: Hybrid, Mbale
- Basıldığı Ülke: Uganda
- Anahtar Kelimeler: CNN, deep learning, hyperparameter optimization, kidney stone classification, NAS, ranking stabilization epoch
- Yozgat Bozok Üniversitesi Adresli: Evet
Özet
Kidney stone detection from computed tomography (CT) images is an important task in medical image analysis, where both classification accuracy and computational efficiency are critical. This study investigates the training dynamics of convolutional neural networks (CNNs) for kidney stone classification and introduces the concept of Ranking Stabilization Epoch (RSE), defined as the earliest epoch at which relative model rankings become reliably consistent with the final ranking. A baseline CNN architecture was systematically modified to generate 10 CNN variants differing in filter configurations, kernel sizes, activation functions, pooling strategies, and regularization techniques. All models were trained and evaluated under identical experimental conditions using a publicly available kidney stone CT dataset. Experimental results showed that full ranking stabilization was achieved at epoch 38 using a predefined Spearman correlation threshold relative to the final 50-epoch ranking. Sensitivity analysis performed under different correlation thresholds demonstrated the robustness of the proposed RSE framework. In contrast to full ranking stabilization, the identity of the best-performing model stabilized considerably earlier, with the top-ranked CNN model maintaining its leading position from the early stages of training onward. The final selected model achieved approximately 97% test accuracy and 98% F1-score. These findings indicate that the proposed framework enables reliable early identification of the best-performing model before complete ranking convergence is achieved. Unlike traditional early stopping or performance prediction methods, the proposed approach explicitly focuses on ranking stabilization during training. Moreover, by estimating the minimum number of epochs required for reliable model comparison, the proposed framework provides a practical and cost-effective strategy for CNN model selection and neural architecture search applications.