Browsing by Author "Acici, Koray"

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • Article
    Balancing Accuracy and Efficiency in Fine-Grained Image Classification: A Systematic Comparison of CNN, Vision Transformer, and Hybrid Architectures
    (Springer Wien, 2026) Eraslan, Emre Cem; Salehi, Hakan; Guzel, Mehmet Serdar; Oztas, Demir Mesut; Ekinci, Fatih; Acici, Koray; Demir, Ozge
    Fine-grained image classification remains a challenging problem in computer vision due to high inter-class visual similarity and pronounced intra-class variability. In this study, a systematic and architecture-centric comparison of convolutional neural networks (CNNs), Vision Transformer (ViT) models, and hybrid architectures is presented using the Oxford 102 Flowers dataset, which comprises 8189 images across 102 visually similar categories. Rather than proposing a new architecture or training strategy, this work focuses on a controlled and comparative analysis of widely adopted deep learning paradigms under a unified experimental framework. All models are evaluated under a strictly controlled experimental protocol with fixed data splits, identical preprocessing, and unified optimization settings to reduce experimental confounding factors and enable architecture-level analysis. The experimental results show that large-scale Vision Transformers achieve the highest absolute classification performance within this benchmark setting, with ViT-L/16 attaining an accuracy of 99.37% and an F1-score of 99.27%. However, this performance is accompanied by substantial computational cost, including 303.4 million parameters and an inference time of 27.4 ms per image. In contrast, hybrid architectures exhibit a more favorable accuracy-efficiency trade-off under the same experimental conditions. Notably, ConvNeXt-V2-Tiny achieves 98.13% accuracy and a 98.05% F1-score while requiring only 27.9 million parameters and an inference time below 1 ms, achieving competitive performance with significantly lower computational complexity. Conventional CNN architectures demonstrate comparatively lower performance within this specific transfer learning configuration, with average accuracy remaining below 90%, suggesting potential limitations in capturing highly subtle inter-class variations under the evaluated setup rather than indicating general architectural inadequacy. The findings indicate that the highest classification accuracy does not necessarily correspond to the most practically efficient solution, particularly in deployment-constrained environments. Importantly, the conclusions of this study are restricted to the Oxford 102 Flowers dataset under transfer learning and standardized input resolution, and should not be interpreted as universal claims of architectural superiority. By jointly analyzing accuracy, model size, and inference latency, this work provides a structured and decision-oriented assessment of architectural trade-offs for fine-grained image classification. The results offer practical insights for selecting appropriate deep learning architectures under varying computational constraints and contribute to a more nuanced understanding of how convolutional, transformer-based, and hybrid representations behave within a controlled fine-grained benchmark setting.
  • Article
    Explainable Brain Tumor Detection in Skull Base CT Using Continuous Neural Representations: Comparative Evaluation of Neural Fields, Neural Operators, and Vision Transformers
    (Pergamon-Elsevier Science Ltd, 2026) Dagli, Ozlem; Guzel, Mehmet Serdar; Kirik, Berk; Ekinci, Fatih; Acici, Koray; Demir, Ozge; Akarkamci, Hilal
    Reliable brain tumor detection in CT remains challenging due to low soft-tissue contrast, skull base complexity, and imaging artifacts. This study evaluates whether continuous neural representation learning improves classification performance and interpretability for CT-based brain tumor detection. We performed a comparative evaluation of coordinate-based neural models, neural operators, and transformer-based vision models using a skull base CT dataset from Gazi University Faculty of Medicine. The dataset includes 200 patients and 40,000 slices (12,000 tumor-positive, 28,000 tumor-negative). An automated multi-stage slice selection framework was developed to extract three anatomically consistent skull base slices per patient using similarity-based ranking and anatomical validation. Ten models were evaluated, including Neural Fields, Implicit Neural Representations, DeepONet, HyperNetworks, FunCNets, Rational Neural Networks, Fourier Neural Networks, and ViT-Base/ Small/Large. Neural Fields achieved the best performance with 98.90% accuracy, 98.90% sensitivity, 98.90% specificity, 0.978 MCC, and 0.990 AUC. FunCNets ranked second overall while DeepONet achieved the highest sensitivity but lower overall balance. Transformer-based models underperformed compared to neural representation approaches. Statistical analysis confirmed significant performance differences with Neural Fields significantly outperforming competing methods. LIME analysis showed that Neural Fields focused on clinically relevant lesion regions while reducing influence from skull base artifacts. Continuous neural representations, particularly Neural Fields, provide superior accuracy and robustness for CT-based brain tumor classification compared to transformer and operator-based models. The findings demonstrate that neural field-based learning offers both high diagnostic performance and improved interpretability, supporting its potential for clinical decision-support in challenging skull base CT analysis.