Akademik Veri Yönetim Sistemi
5th International Conference on Pioneer and Innovative Studies – ICPIS 2026, Konya, Türkiye, 8 - 09 Haziran 2026, ss.20, (Özet Bildiri)
Abstract – Accurate and consistent diagnosis brain CT scan images plays a crucial role for decision support systems. This study, three different deep learning schemes are compared for stroke diagnostic. This study utilized a dataset from the Teknofest 2021 competition containing 4427 stroke-free, 1130 ischemia and 1093 hemorrhage classes. To control dataset imbalance, the Categorical Focal Cross Entropy loss function was prefered in the training process. Data preprocessing step, black borders were removed, converted the images to 512x512 pixels. 5-fold cross-validation strategy was applied for it to be generalized of the models accuracy. Each model was trained with the same hyperparameter settings to create a fair comparison environment. Performance metrics evaluated Accuracy, Precision, Recall, AUC, ROC Curve and Macro F1. Model diagnoses analyzed compare heat maps generated with Grad-CAM, an Explainable Artificial Intelligence tool, to expert markings in the dataset. According to performance metrics, ViT model achieved 86.59% Macro F1 and 89.92% Accuracy, while EfficientNetV2 model also exhibited high performance with 86.42% Macro F1 and 90.23% Accuracy scores. ConvNeXt failed to generalize and classified all images as hemorrhage, resulting in the lowest performance. The most critical finding of the study has been revealed with Grad-CAM analysis. Although ViT achieved high scores on the metrics, an analysis of the heatmaps revealed that there were different focal points across the entire image. EfficientNetV2 showed a high degree of similarity to the expert marking and heatmaps generated. The results of the study have shown that reliability in disease diagnosis processes does not depend solely on numerical metrics. Models that appear numerically successful may focus on different features in the image for disease diagnosis. As a result, EfficientNetV2 has been identified as the most reliable and appropriate choice in terms of accuracy and consistency.