NKNida Kumbasar

Nida Kumbasar

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

Nida Kumbasar has a total of 13 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • PATCH-BASED TOOTH SEGMENTATION ON CHILDREN'S DENTAL PANORAMIC RADIOGRAPHS

    by Nida Kumbasar, Emin Argun Oral, Yücel Özbek

    Dental Panoramic Radiography (DPR) is a widely utilized imaging modality for assessing oral health. However, accurate tooth segmentation in pediatric DPRs remains a significant challenge due to the variable morphology of deciduous teeth and the inherent complexity of overlapping anatomical structures in 2D projections. This study was conducted to enhance segmentation accuracy in pediatric DPRs by employing a patch-based deep learning (DL) approach. The models were trained using the relatively recent CDPR dataset, which consists of 193 pediatric DPRs and their corresponding ground truth masks. To investigate the optimal input size, the images were divided into 2, 4, 8, 16, and 32 patches, in addition to the full-resolution images. Five prominent DL models—PSPNet, DeepLabv3+, R2 U-Net, U-Net, and U-Net++—were trained and evaluated across these different patch configurations. A subject-based patch reconstruction method was subsequently implemented to rebuild the full-size segmented images for comprehensive evaluation. Experimental results indicated that the optimal segmentation performance for PSPNet, DeepLabv3+, and U-Net was achieved with the 8-patch approach, while R2 U-Net performed best with the 16-patch configuration. Notably, the U-Net++ model using a 2-patch approach outperformed all other models and configurations, achieving the highest Dice score of 0.9222. These findings demonstrate that patch-based training significantly improves segmentation performance in pediatric DPRs, which is a particularly valuable outcome under limited data conditions. This novel approach offers promising results for clinical integration, effectively supporting pediatric dental care by enhancing segmentation accuracy and overall workflow efficiency.

  • Open Access

    Fusion based collective intelligence segmentation using panoramic radiograph images

    by Nida Kumbasar, Emin Argun Oral, Ibrahim Yucel Ozbek

    In recent years, there has been an increasing body of research highlighting the close relationship between the diagnosis of osteoporosis and the lower jaw. The study of the jawbone in dentistry has gained increasing popularity and has become a focus for artificial intelligence applications. This study, employing a collective intelligence modality, achieved pixel-level fusion by integrating three recently prominent segmentation methods for mandible and maxillomandibular segmentation. Two publicly available third-party datasets were utilized for mandible and maxillomandibular segmentation. The collective intelligence modality underwent analysis on both hard label (0 or 1) and soft label ([0-1]) at the pixel level through decision fusion. Ultimately, it was observed that the decision fusion process with soft labels surpassed the performance of both single modality and hard label decision fusion. In ablation studies, the threshold parameter for transitioning from soft label to hard label in binary segmentation was determined by assessing the model’s dice score performance using the grid search approach. As a result of the experimental studies, the dice score of the Soft Label Decision Fusion by mean model ( $$SLDF\underline mean$$ ), SLDFNet, which achieved the highest performance on both datasets, was 0.9821 for the mandible and 0.9782 for the maxillomandibular region, respectively. The binary segmentation threshold parameter of SLDFNet in both datasets was determined to be 0.4 through experiments. According to the results, it seems that the proposed method provides state-of-the-art outcomes. The robustness of the proposed method has been demonstrated on two datasets, each with relatively few and large amounts of data compared to the other.


  • Open Access

    Winter Classification of Mandibular Impacted Third Molars Based On Convolutional Neural Networks

    by Mustafa Taha Güller, Nida Kumbasar, Ozkan Miloglu, İbRahim Yücel Özbek, Emin Argun Oral

    Objective: Winter classification (WC) is used in the radiographic evaluation of mandibular impacted third molars (MITM) before extraction. In our study, we investigated the classification performance of panoramic radiographs (PRs) using different versions of two convolutional networks (CNN).Methods: The analysis was performed using three different YOLO-v7 and five different YOLO-v8 CNN architectures for the WC of 716 MITM teeth in 532 PRs included in the dataset. The localization of the second and third molars on PR images was determined, and the diagnostic performance of WC in this area was measured. Precision, recall, and mean average precision (mAP) were statistically evaluated for each model.Results: For both architectures, the highest performance was obtained for horizontal classification, with a value of 0.917 for the mAP metric, while the lowest performance was found to be 0.799 for vertical classification. Looking at the mAP metric values for all classes in the study, YOLO-v8-m performed better than YOLO-v7, with a difference of 2.7%, resulting in an overall mAP of 0.838 for YOLO-v7 and 0.865 for YOLO-v8. For both YOLO-v7 and YOLO-v8, the mid-depth network performed better than the other sub-models.Conclusions: This is the first study in which WC is performed using YOLO-v7 and YOLO-v8 models. In line with our results, CNN models for WC in PR are promising. In future studies, better results can be obtained by increasing the dataset size, using images from different centers and developing CNN architectures. Thus, clinical use of artificial intelligence architectures may become widespread.Keywords: Convolutional neural network, object detection, panoramic radiography, third molar, winter


  • Drone Classification with a Hybrid Deep Learning Approach Based on Mel-Spectrogram Representation

    by Nida Kumbasar, Rabiye Kılıç, Emin Argun Oral, Yücel Özbek

    With technological advancements, the use of drones has become increasingly widespread in both civilian and military sectors in recent years. There is a need for technologies that can detect and identify the presence, type, or flight mode of a drone with remote sensing signals in situations that pose a security threat. This study is based on the classification of radio frequency (RF) signals from various drones under different flight modes using Mel-spectrogram representations. Within the scope of the study, the VGG19 deep learning model is used as both a classifier and a feature extractor for the SVM classifier. On the other hand, the study compares the performance of RF signals in low and high frequency bands separately and in concatenated versions. In the results obtained in the experimental studies, the VGG19+SVM hybrid model showed the highest performance over the Mel-spectrogram of the concatenated low and high (L+H) frequencies. The accuracy performances were 100% in the 2-Class problem where drone presence was detected (drone -no drone), 90.78% in the 4-Class problem where drone types were classification (Bebop-AR-Phantom-no drone), and 86.6% in the 10-Class problem where drone modes were obtained.


  • Open Access

    Comparison of VT-based and CNN-based Models on Teeth Segmentation

    by Şilan Fidan Vural, Nida Kumbasar

    Semantic segmentation is a crucial task in computer vision with a wide array of applications across various fields, especially in medical imaging. One of the most important applications of semantic segmentation is in the field of dentistry, where teeth segmentation plays a significant role in diagnosing and treating oral health issues. Accurate segmentation of teeth in dental images is vital for detecting abnormalities, planning treatments, and monitoring the progress of dental procedures. In this paper, a comprehensive comparative analysis is presented, focusing on the use of Convolutional Neural Network (CNN)-based and Vision Transformer (VT)-based models for image segmentation within the context of dentistry. The paper presents a comparison of eight different models, contributing to the literature on dental image segmentation and showcasing practical applications in clinical dental settings. The research presented in this study uses several state-of-the-art segmentation models, namely U-Net, LinkNet, and Swin U-Net, along with different backbones to perform teeth segmentation on publicly available two datasets: one representing adults and the other children. The experiments were conducted to determine which models and backbones provided the best segmentation performance for each dataset. The study also emphasizes that the segmentation modeling process should be handled separately since the alignment of child and adult teeth is different. The U-Net model with the ResNet101 backbone achieved the best performance on the adults dataset, while for the children dataset, the U-Net model with the same ResNet101 backbone also demonstrated superior results. The highest Dice scores obtained were 0.9543 for the adults dataset and 0.9019 for the children dataset, indicating the effectiveness of these models in accurately segmenting teeth. The findings from this research demonstrate the potential of deep learning techniques in improving the accuracy and efficiency of dental diagnosis and treatment planning. Codes used throughout the study will be publicly available at https://github.com/FidanVural/Teeth-Segmentation-in- Panoramic-Radiography/tree/main


  • Open Access

    Comparison of Deep Learning Algorithms for Image Segmentation on Satellite Images

    by Hüseyin Acemli, Nida Kumbasar

    Recent advancements in deep learning have significantly contributed to the development of high spatial resolution (HSR) land cover mapping. However, the distinct geographic patterns between urban and rural areas have limited the generalizability of deep learning algorithms across these domains. To address this challenge, separate datasets for rural and urban environments have been proposed in the literature, aiming to achieve more reliable results in real-world applications. In this study, we utilize the publicly available LoveDA HSR dataset for model and parameter comparison. Experiments were conducted on two distinct scenarios: rural and urban areas. The combination of the Adam optimizer, Dice loss function, and UNet++ architecture exhibited the highest performance in both datasets. A weighted average of this combination, based on the number of test samples, was calculated for both groups, yielding a final performance score of 62.14% in terms of mean Intersection over Union (IoU).