RARamazan Savas Aygün

Ramazan Savas Aygün

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Ramazan Savas Aygün has a total of 110 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Parameter-Efficient Hybrid Quantum-Classical Learning for Diabetic Retinopathy Grading: Development and External Validation Study (Preprint)

    by Md Rashik Shahriar Akash, Sait Suer, K M Ferdous, Ramazan Aygun, Mahmut Karakaya

    BACKGROUND Diabetic retinopathy (DR) grading becomes more difficult when models are applied to images from new clinical sources. Differences in cameras, image quality, and patient populations can reduce model performance after transfer. Hybrid quantum-classical models can reduce the number of parameters that need to be trained, but previous quantum studies for DR have been generally evaluated on a single dataset. The key question is whether hybrid quantum-classical learning can be useful for cross-domain DR grading performance across new datasets while reducing task-specific training from millions of parameters to thousands. OBJECTIVE We developed a parameter-efficient hybrid quantum-classical model for DR grading and evaluated how well it transfers to external datasets. We compared it with fully fine-tuned classification and regression models and examined overall performance, grade-level errors, and confidence calibration. METHODS We trained all models on APTOS 2019 and tested them without further tuning on MESSIDOR-2 (n=1,744), IDRiD (n=516), and UoA (n=200). Our Hybrid-QC model uses a frozen ConvNeXt-Tiny encoder followed by a small classical projection, an 8-qubit variational quantum circuit, and a final classification layer. Hybrid-QC updates 25,093 parameters, compared with about 28.05 million trainable parameters in the fully fine-tuned ConvNeXt-Tiny classifier (ConvNeXt-Cls) and regression model (ConvNeXt-Reg), approximately 1,118 times fewer. We first evaluated no-DR versus advanced-DR classification and then the more difficult five-class grading. RESULTS For no-DR versus advanced-DR classification, Hybrid-QC achieved an area under the receiver operating characteristic curve (AUROC) of 0.986 on MESSIDOR-2, 0.989 on IDRiD, and 0.995 on UoA. All three models achieved AUROC above 0.98 on every external dataset, and no two models differed by more than 0.01. In five-class grading, Hybrid-QC achieved quadratic weighted kappa (QWK) values close to those of ConvNeXt-Cls on MESSIDOR-2 (0.608 vs 0.618) and IDRiD (0.719 vs 0.717). ConvNeXt-Reg achieved the highest QWK and the lowest mean absolute error on all three external datasets. Hybrid-QC achieved the highest Grade-1 F1 on MESSIDOR-2, IDRiD, and UoA (0.186, 0.162, and 0.400) and predicted fewer Grade-1 cases as Grade 0 than ConvNeXt-Cls. It also had a lower expected calibration error than ConvNeXt-Cls on MESSIDOR-2 (0.047 vs 0.296), IDRiD (0.141 vs 0.373), and UoA (0.100 vs 0.216), and remained better calibrated even after temperature scaling of ConvNeXt-Cls. CONCLUSIONS Hybrid-QC reduces the trainable parameter requirement from about 28 million to about 25 thousand while achieving binary AUROC within 0.01 of the fully fine-tuned models on all three external datasets and five-class QWK close to ConvNeXt-Cls on MESSIDOR-2 and IDRiD. ConvNeXt-Reg remained stronger for overall ordinal grading, while Hybrid-QC achieved the highest Grade-1 F1 on all three external datasets and lower calibration error than ConvNeXt-Cls after domain shift. These results support Hybrid-QC as a parameter-efficient approach to cross-domain DR grading.

  • Why is Data Misclassified? Quantifying Data Difficulty for Machine Learning Models with Dual Probability Difficulty Measure

    by Ayomide Afolabi, Ramazan Aygun, Truong X. Tran

    Despite the high accuracy of machine learning models, misclassification of data may raise issues regarding the trustworthiness of these models. There are a number of reasons for misclassification, such as model complexity and the difference between the distributions of training data and evaluation data. In this paper, we investigate whether specific data could be hard to classify correctly based on the training dataset or not. The development of measures to quantify the difficulty of data provides insight into assessing the potential performance of such models. However, most of the existing data difficulty measures, which are typically model-based, suffer from various model biases, while the few existing model-agnostic measures do not yield consistent results across datasets and machine learning models. In this paper, we propose a novel model-agnostic measure named Dual Probability Difficulty Measure. This measure considers the probability of the data instance label and the probability of the most frequent label that does not share the same label as the data instance in the K-nearest neighbors to assess the data difficulty. We conducted experiments to evaluate the accuracy of machine learning models with respect to different data difficulty levels and utilized the area under the curve to assess the effectiveness of our proposed method compared to other methods. Our experimental results on five diverse datasets across various machine learning models show that our Dual Probability Difficulty Measure has a correlation with model performance and identifies data likely to be misclassified.


  • A Convolutional Neural Network Ensemble for Video Source Camera Forensics

    by Maryna Veksler, Ramazan S. Aygun, Kemal Akkaya, Sitharama S. Iyengar

    The advancement of IoT technologies influenced a substantial increase in the usage of multimedia devices. Consequently, forensic specialists began to experience a vast load of video and image data during the investigative procedures. This triggered a need to ensure the integrity of multimedia data and verify its source origin for digital forensics processes. In this paper, we address the problem of identifying the video source camera of the video data acquired by the investigators. We develop a novel convolutional neural networks (CNNs) ensemble framework to identify the video source camera. In our method, we analyze the video data using patches extracted from I-frames quadrants (i.e., non-overlapping squares) using independent CNNs for each quadrant to achieve location awareness. Experimental results demonstrate that our framework is robust for the same device type classification and outperforms existing deep learning-based techniques.


  • Query-by-Gaming

    by Ramazan S. Aygun, Vineetha Bettaiah

  • Temporal information retrieval using bitwise operators

    by Prasanna Koirala, Ramazan Aygun, Tathagata Mukherjee, Haeyong Chung


  • Open Access

    A Multistage Framework for Detection of Very Small Objects

    by Duleep Rathgamage Don, Ramazan Aygun, Mahmut Karakaya

    Small object detection is one of the most challenging problems in computer vision. Algorithms based on state-of-the-art object detection methods such as R-CNN, SSD, FPN, and YOLO fail to detect objects of very small sizes. In this study, we propose a novel method to detect very small objects, smaller than 8×8 pixels, that appear in a complex background. The proposed method is a multistage framework consisting of an unsupervised algorithm and three separately trained supervised algorithms. The unsupervised algorithm extracts ROIs from a high-resolution image. Then the ROIs are upsampled using SRGAN, and the enhanced ROIs are detected by our two-stage cascade classifier based on two ResNet50 models. The maximum size of the images used for training the proposed framework is 32×32 pixels. The experiments are conducted using rescaled German Traffic Sign Recognition Benchmark dataset (GTSRB) and downsampled German Traffic Sign Detection Benchmark dataset (GTSDB). Unlike MS COCO and DOTA datasets, the resulting GTSDB turns out to be very challenging for any small object detection algorithm due to not only the size of objects of interest but the complex textures of the background as well. Our experimental results show that the proposed method detects small traffic signs with an average precision of 0.332 at the intersection over union of 0.3.


  • Open Access

    Automation of Explainability Auditing for Image Recognition

    by Duleep Rathgamage Don, Jonathan Boardman, Sudhashree Sayenju, Ramazan Aygun, Yifan Zhang, Bill Franks, Sereres Johnston, George Lee, Dan Sullivan, Girish Modgil

    XAI requires artificial intelligence systems to provide explanations for their decisions and actions for review. Nevertheless, for big data systems where decisions are made frequently, it is technically impossible to have an expert monitor every decision. To solve this problem, the authors propose an explainability auditing method for image recognition whether the explanations are relevant for the decision made by a black box model, and involve an expert as needed when explanations are doubtful. The explainability auditing system classifies explanations as weak or satisfactory using a local explainability model by analyzing the image segments that impacted the decision. This version of the proposed method uses LIME to generate the local explanations as superpixels. Then a bag of image patches is extracted from the superpixels to determine their texture and evaluate the local explanations. Using a rooftop image dataset, the authors show that 95.7% of the cases to be audited can be detected by the proposed method.


  • Spectra-based Classification of Audiovisual and Visuospatial Face-name Associative Memories using EEG

    by Femi William, Feng Zhu, Ramazan Aygun, Mattie Ponter

    In social interactions, we remember faces and associate names that we hear to the faces. Existing research has studied single-item memory tasks such as names, face, colors, shapes, and classified their associations using spectral analysis, event-related potentials(ERPs), and non-ensemble machine learning techniques. However, the impact of multiple stimuli presentation modalities in the associative memory of face-name stimuli has not been extensively investigated. In this research, we conducted two experiments on audiovisual face-name and visuospatial face-name memory tasks. We recorded EEG data from 15 healthy participants and extracted the average oscillatory band power of the alpha, theta, beta, and gamma bands and the signal entropy. Adaptive Logistic boosting was applied for classification, and t-test statistical analyses were applied to investigate participants’ behavioral performance. We found that people did better on visuospatial face-name memory association than audiovisual face-name memory association. Adaptive boosting performed best in comparison to other existing algorithms with the average classification accuracy of 76.37% at the parietal electrode P 7 .