FSM Vakıf Üniversitesi Araştırma ve Akademik Performans Sistemi
DSpace@FSM, FSM Vakıf Üniversitesi’nin bilimsel araştırma ve akademik performansını izleme, analiz etme ve raporlama süreçlerini tek çatı altında buluşturan bütünleşik bilgi sistemidir.
Güncel Gönderiler
Öğe Türü: Öğe , Quantization Techniques for Resource-Efficient Medical Image Segmentation Across Diverse Modalities(ELECO, 2025) Göksu, Tuğçe; Aydın, Musa; Kuş, ZekiDeep learning–based medical image segmentation models achieve high accuracy but are often impractical for clinical deployment due to their computational and memory demands. This study evaluates Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) on a U-Net architecture across seven MedSegBench datasets. Both approaches reduced model size from ~124 MB to ~31 MB (3.93× compression) and achieved 2.4–3.7× faster inference on CPU. PTQ delivered these gains with minimal accuracy loss, maintaining F1 scores nearly identical to full-precision baselines (e.g., 0.911 vs. 0.911 on Bbbc010, 0.958 vs. 0.958 on Pandental). QAT, while requiring up to 61k seconds of training, outperformed FP32 in challenging datasets, improving F1 by 25.7% on Nuclei and 3.7% on UWSkinCancer. These results demonstrate that PTQ is optimal for rapid, resource-limited deployment, whereas QAT ensures superior robustness and accuracy. Quantization thus provides a practical pathway for efficient and reliable medical image segmentation on clinical systems.Öğe Türü: Öğe , Multi-Objective Surrogate-Assisted Neural Architecture Search for Digital Pathology: A Tree-Based Regression Approach(ELECO, 2025) Akçelik, Zeliha Kaya; Dik, Sümeyye Zülal; Kuş, Zeki; Aydın, MusaDigital pathology enables histopathological slide digitization, becoming crucial in modern medicine. While deep learning models excel at classification, their success requires carefully designed architectures, making manual design expensive. Neural Architecture Search (NAS) automates this process, but high evaluation costs limit these methods. Surrogateassisted NAS addresses this by predicting architecture performance without full training. We propose a multiobjective surrogate-assisted NAS method for digital pathology whole slide image classification. Our approach simultaneously minimizes classification error and parameter count using Path-based Multi-Objective Differential Evolution with opposition-based learning. Unlike CNN-based surrogates, we employ tree-based regression models (CatBoost, LightGBM, XGBoost). EBHI dataset experiments show model complexity directly impacts accuracy and inference time. The CatBoost-High model achieved the highest accuracy (98.73%), while XGBoost-Medium provided optimal trade-off (98.46% accuracy, 250s). Results demonstrate our method achieves high accuracy while significantly reducing computational cost, making it promising for digital pathology applications.Öğe Türü: Öğe , Distilling Knowledge or Transferring Weights? An Experimental Perspective on Classifiers(ELECO, 2025) Öğ, Merve; Yıldızlı, Beyza; Kuş, Zeki; Aydın, MusaThis study presents a systematic comparative analysis of knowledge distillation and transfer learning methodologies applied to image classification on the CIFAR-10 dataset. Using ResNet-18 architectures as the baseline, we investigate the trade-offs between model complexity, computational efficiency, and classification performance under various optimization strategies. Results demonstrate that knowledge distillation consistently outperforms transfer learning across all tested configurations. Most notably, a lightweight ResNet- 18 student model (2.84M parameters) guided by a ResNet-18 teacher achieved 89.03% accuracy, significantly exceeding transfer learning's 86.36% maximum accuracy despite using only 25% of the parameters. This improvement changes how we optimize models. It shows that using soft targets for knowledge transfer can beat the usual trade-off between a model's size and how well it performs. This makes it useful for places with limited resources.Öğe Türü: Öğe , A Systematic Comparison of Text and Image Encoders for Visual Question Answering: From RNN to LLM-Based Representations(ELECO, 2025) Dik, Sümeyye Zülal; Hoşavcı, Reyhan; Akçelik, Zeliha Kaya; Kuş, Zeki; Aydın, MusaThis study presents a systematic comparative analysis of text and image encoder combinations for Visual Question Answering (VQA) using the EasyVQA dataset. We evaluate six text encoders (ELMo, BERT, RoBERTa, T5, SBERT, LLM2Vec) paired with three image encoders (ResNet-50, DenseNet-121, ViT) under both frozen and fine-tuned training scenarios. Our two-branch architecture processes images and questions separately before concatenating embeddings for classification. Results demonstrate significant performance variations between training strategies, with finetuning improving average accuracy from 93% to 96%. LLM2Vec achieved perfect performance (100% accuracy) with DenseNet-121 in frozen mode, while BERT and RoBERTa showed remarkable improvements through finetuning, reaching perfect scores with multiple image encoders. DenseNet-121 proved most stable across configurations. These findings reveal that modern LLM-based encoders excel with minimal adaptation, while traditional Transformer models benefit substantially from task-specific fine-tuning, providing crucial guidance for multimodal system design.Öğe Türü: Öğe , Time-Varying Inflation Co-Movement: Dynamic Factor Model with Global, Group, and Country-Specific Factors(Elsevier, 2026) Tunç, Ahmet; Nazlıoğlu, Şaban; Yücel, Ali GökhanThis study examines global inflation co-movement by employing a dynamic factor model with time-varying parameters and stochastic volatility to extract global, group, and country-specific factors across 86 developed and developing economies from 1971 to 2023. The results reveal that (i) the global factor has gradually receded as the primary source of inflation variation, giving way to more group- or country-specific dynamics—though brief episodes of global synchronization reemerge during major crises; (ii) the group factor dominates inflation variation in developed countries; and (iii) the country-specific factor remains the main driver of inflation in developing countries.
