Ramazan Terzi

Early and accurate diagnosis of breast cancer is critical for reducing mortality and improving patient outcomes, as it enables timely intervention and more effective treatment planning. In this context, developing reliable and high-performance automated systems for mammography analysis is of great clinical and societal importance. In this study, the classification efficiency of MedSigLIP, one of the most recent technologies in the medical domain, was evaluated on mammography images. Within this scope, both its feature extraction capability and fine-tuning model training competence were assessed, and comprehensive experiments were conducted. For the feature extraction task, features were extracted using MedSigLIP and SigLIP2, the state-of-the-art vision encoder, and their classification performances were reported using seven different machine learning methods. In addition, for the fine-tuning model training task, MedSigLIP was compared with six different convolutional neural network architectures, and the results were reported. According to the obtained results, MedSigLIP achieved the best performance in both tasks, producing 73% F1 score in the MLO view and 83% F1 score in the CC view for feature extraction, and 76% F1 score in the MLO view and 89% F1 score in the CC view for the fine-tuning task.

Artificial intelligence applications used to monitor plant health and growth, especially for commercially valuable crops, are important for both scientific research and economic value. Despite the availability of numerous datasets, most of them are designed primarily for disease classification. In contrast, datasets annotated at the leaf instance level remain limited. This limitation constrains many studies to classification-based approaches, restricting the development of more advanced tasks such as instance-level segmentation. In this study, leaves from 13 different plant datasets, most of which were originally for disease classification, were segmented at the instance level and introduced as a new dataset named LeavesBank. Within LeavesBank, 32,232 unique images containing real backgrounds, gathered using different devices and across various geographic regions, were combined. Subsequently, an active learning–based annotation pipeline was implemented, through which a total of 221,427 leaf instances were segmented, including 198,489 primary and 22,938 secondary annotations. The effectiveness of the LeavesBank was evaluated using four public datasets, through both zero-shot evaluation and fine-tuning experiments. According to the experimental results, LeavesBank-pretrained models achieved 10%, 5%, and 1% higher mAP50 scores on the Komatsuna, GrowliFlower and SoyCotton datasets, respectively, compared with the most recent foundation model, SAM3. Only on the Potato Leaf dataset did SAM3 outperform the LeavesBank-pretrained models, achieving a 9% higher mAP50 score. Furthermore, in experiments, the proposed approach demonstrated 1%–24% improvements in mAP50 performance according to the literature. The LeavesBank dataset is publicly available at: https://www.kaggle.com/datasets/aslhanyldrm/leavesbank-dataset

Ensuring plant sustainability is critically important across numerous domains. Specifically, the detection and segmentation of leaves are essential for tasks such as identifying plant diseases, monitoring plant growth, and determining plant phenotypes. However, the limited diversity in both data and species within existing datasets prepared for instance segmentation tasks often leads to the development of models with poor generalization capabilities and significant biases. This study introduces GenYOLO-Leaf, a data-centric, open-source framework designed to address these limitations. GenYOLO-Leaf facilitates instance-based leaf segmentation with improved generalization capabilities using different data sets with enriched label information by extracting approximately 145K leaf instances. It can also serve as a valuable resource for transfer learning across various segmentation tasks. The developed framework underwent zero-shot evaluation using a total of eight distinct datasets: four for instance segmentation and four for semantic segmentation. Experimental results indicate that the framework achieved mAP scores ranging from 62 to 84% on instance segmentation datasets, while producing mean IoU scores between 86 and 99% on semantic segmentation datasets. The GenYOLO-Leaf framework that includes model weights for YOLOv11 and YOLOv8 is publicly available at https://github.com/aaslihanyildirim/GenYOLO-Leaf .

Evaluating artificial intelligence for a focal nodular hyperplasia diagnosis using magnetic resonance imaging: preliminary findings - Diagnostic and Interventional Radiology

This paper proposes ensemble strategies for the deep learning object detection models carried out by combining the variants of a model and different models to enhance the anatomical and pathological object detection performance in brain MRI. In this study, with the help of the novel Gazi Brains 2020 dataset, five different anatomical parts and one pathological part that can be observed in brain MRI were identified, such as the region of interest, eye, optic nerves, lateral ventricles, third ventricle, and a whole tumor. Firstly, comprehensive benchmarking of the nine state-of-the-art object detection models was carried out to determine the capabilities of the models in detecting the anatomical and pathological parts. Then, four different ensemble strategies for nine object detectors were applied to boost the detection performance using the bounding box fusion technique. The ensemble of individual model variants increased the anatomical and pathological object detection performance by up to 10% in terms of the mean average precision (mAP). In addition, considering the class-based average precision (AP) value of the anatomical parts, an up to 18% AP improvement was achieved. Similarly, the ensemble strategy of the best different models outperformed the best individual model by 3.3% mAP. Additionally, while an up to 7% better FAUC, which is the area under the TPR vs. FPPI curve, was achieved on the Gazi Brains 2020 dataset, a 2% better FAUC score was obtained on the BraTS 2020 dataset. The proposed ensemble strategies were found to be much more efficient in finding the anatomical and pathological parts with a small number of anatomic objects, such as the optic nerve and third ventricle, and producing higher TPR values, especially at low FPPI values, compared to the best individual methods.

Age estimation from brain MRI has proved to be considerably helpful in early diagnosis of diseases such as Alzheimer’s and Parkinson’s. In this study, curriculum learning effect on age estimation models was measured using a brain MRI dataset consisting of normal and anomaly data. Three different strategies were selected and compared using 3D Convolutional Neural Networks as the Deep Learning architecture. The strategies were as follows: (1) model training performed only on normal data, (2) model training performed on the entire dataset, (3) model training performed on normal data first and then further training on the entire dataset as per curriculum learning. The results showed that curriculum learning improved results by 20 % compared to traditional training strategies. These results suggested that in age estimation tasks datasets consisting of anomaly data could also be utilized to improve performance.

Bu çalışmada, filo yönetim sistemlerinin eksikleri hem büyük veri hem de sürücü davranışı bakış açılarıyla incelenmiş, gerçek veriler üzerinde sürücü/sürüş davranışları analiz edilmiş bunun sonucunda veri analizlerini gerçekleştirmek için büyük veri tabanlı yeni modeller önerilmiştir. Büyük veri üzerinde 6 farklı senaryo ile filodaki sürücülerin davranış farklılıkları, çeşitli lokasyonlardaki davranışları ve spesifik noktalardaki davranışları büyük veri temelli yaklaşımlar ile tespit edilmiş yeni modeller geliştirilmiştir. Elde edilen sonuçlardan, özellikle %50 üzerinde hız limiti aşan sürücüler arasında belirli sürücülerin diğer sürücülere göre %30 oranında bu ihlallerde payı olduğu, ortalama hızları aynı olsa bile hız ihlalleri bakımından sürücüler arasında 6 kat farklılık olabileceği, benzer şekilde hız ihlal sayıları aynı olsa bile ihlal sürelerinde 2 kata yakın fark olabileceği, mevsimlere göre hız ihlal sayısının en fazla yaz mevsiminde fakat ihlal süresinin en fazla sonbaharda olduğu, Ankara ili içinde hız limitlerini aşmaya olanak sağlayan yolların ilçe bazında %23,6 oranıyla Yenimahalle, mahalle bazında %4,62 oranı ile Saray, şehirlerarası yol bazında %6,85 oranı ile Eskişehir yolunun, şehir içi yollarda ise %2,74 oranı ise Anadolu Bulvarı olduğu, son olarak ise radar noktaların yaklaşık 1 kilometre öncesi ve sonrası hız ihlal sayılarına göre 3 farklı radar noktası incelemesinde %300’e yakın oranlarda farklılıkların oluştuğu tespit edilmiştir. Sonuç olarak büyük veri analitiği kullanarak filoların sürücü/sürüş davranış kapsamında daha etkili kullanılabileceği ve daha kolay yönetilebileceği, gerek maliyet gerekse iş gücü kaybının önlenmesi için bu modellerden faydalanılabileceği, Ankara ili için yapılan bu analizlerin diğer iller için de kullanılabileceği, radar noktalarında yapılan analizler gibi farklı değerler üretilebileceği değerlendirilmektedir.