Performance and Efficiency Trade-Off Analysis of Modern CNN Architectures for Automated Stem Cell Colony Quality Assessment
Hatice Hilal Gurleyen, Nejat Yumusak
article (Uluslararası, Hakemli) | Yazar(lar): Hatice Hilal Gurleyen, Nejat Yumusak | Yıl: 2026 | Dergi/Kurum: IEEE ACCESS | İndeks: WOS.SCI | DOI: 10.1109/access.2026.3692455 | Anahtar Kelimeler: Architecture, Calibration, Computer architecture, Modeling, Accuracy, Licenses, Stem cells, Training, Convolutional neural n
Quality assessment of stem cell colonies plays a critical role in regenerative medicine and cell-based therapies. Performing this assessment manually by experts is both time-consuming and prone to inter-observer variability. In addition, the limited availability of high-quality labeled stem cell colony images increases the risk of overfitting in data-driven approaches and complicates model generalization. In this study, a systematic comparison of multiple CNN architectures is presented for the binary classification problem of stem cell colony images as good or bad, within a multi-dimensional evaluation framework incorporating performance, calibration, and efficiency. On an open-access dataset previously reported in the literature with an accuracy of approximately 89%, eight modern CNN architectures, including VGG13, VGG16, ResNet50V2, DenseNet169, InceptionV3, MobileNetV3-Small, EfficientNet-B0, and EfficientNetV2-S, were evaluated using standard preprocessing and data augmentation strategies under a five-fold cross-validation protocol. The models were analyzed using conventional performance metrics such as accuracy, F1-score, ROC-AUC, and average precision, as well as calibration metrics including the Brier score, expected calibration error, and negative log-likelihood. Post-hoc calibration via temperature scaling was applied using a split-validation protocol. In addition, the models were compared in terms of computational cost, including GFLOPs, VRAM consumption, number of parameters, training time, and inference latency, and data efficiency was assessed via learning curve analysis. The results demonstrate that the EfficientNetV2-S model achieves the most balanced performance in terms of accuracy, generalization, and calibration, reaching an accuracy of 95.2%, while the EfficientNet-B0 model provides competitive performance with a substantially lower computational cost. Calibration performance improves notably after post-hoc scaling, highlighting the importance of reliable confidence estimation. These findings emphasize the critical importance of balancing performance, efficiency, and reliability in model architecture selection for data-limited biomedical imaging scenarios and present a practical deep learning framework for the automated quality assessment of stem cell colonies.