Önder Polat

Background: Artificial intelligence (AI) has increasingly been integrated into various fields of medicine, necessitating rigorous oversight by scientists to ensure its appropriate application. This study presents an analysis of microhemorrhages, which are closely associated with significant cerebrovascular events. Researchers have designed a study to investigate the feasibility of utilizing AI in the detection and assessment of microhemorrhages. Methods: The study retrospectively analyzed 108 patients with cerebral microhemorrhagic foci and 108 controls using cranial magnetic resonance imaging (MRI) and susceptibility-weighted sequences, following ethical approval. Deep learning models, including a customized DenseNet for classification and nnU-Net for segmentation, were trained on preprocessed MRI data to detect and localize microhemorrhages efficiently. Results: The study involved a total of 216 patients with and without cerebral microbleeds. There were a total of 146 foci in the 108 patient group. The classification model demonstrated an area under the curve of 80% ± 2.7% and an accuracy of 75% ± 1.7%, though specificity was more limited at 68% ± 4.3%. The segmentation model identified microhemorrhagic regions with a sensitivity of 89.4% ± 1.4% and achieved a Dice similarity coefficient of 62.05% ± 2.3%. Conclusion: Overall, both models demonstrated the potential for detecting and localizing microhemorrhages in 3D MRI data, though the classification model showed moderate specificity. The findings suggest that AI-based approaches may eventually serve as a supportive tool for microhemorrhage detection, provided further refinements are made to enhance detection accuracy and clinical applicability. Cite this article as: Akkaya HE, Polat Ö, Çolakoğlu H, Akpinar Dİ, Kaya E. Deep learning-based detection, classification, and segmentation of cerebral microbleeds. Eurasian J Med. 2026, 58(4), 1414, doi: 10.5152/eurasianjmed.2026.261414.

Aim: To investigate whether MRI-derived texture (radiomic) features combined with machine learning can provide objective diagnostic discrimination between patients with hemifacial spasm (HFS) and healthy controls, and to evaluate the added value of high-resolution 3D T2-SPACE imaging.Materials and Methods: This retrospective study included 35 patients with clinically diagnosed HFS and 36 age- and sex-matched controls. All subjects underwent 3-Tesla MRI with axial T2-weighted imaging (T2WI), and a subgroup of 21 HFS patients and their matched controls also underwent 3D T2-SPACE imaging. Segmentation was performed across axial slices encompassing the brainstem, the entire posterior fossa, and the inner ear, middle ear, and mastoid portions, with the anterior boundary extending to the clivus, using a single manually drawn region of interest (ROI) to capture comprehensive radiomic features. 137 radiomic texture features were extracted using the PyRadiomics platform. Multiple machine learning classifiers, including k-nearest neighbors, Naive Bayes, discriminant analysis, decision trees, and ensemble learning methods, were evaluated. Data were split into 80% training and 20% testing sets, and performance was assessed using accuracy, sensitivity, specificity, and ROC-AUC over 20 randomized repetitions.Results: Among all tested models, the Ensemble Bagged Trees classifier demonstrated the highest overall performance among the tested models. Using T2WI, the model achieved a mean accuracy of 0.71 ± 0.10, sensitivity of 0.72 ± 0.15, specificity of 0.70 ± 0.18, and ROC-AUC of 0.76 ± 0.16. In the subgroup analyzed with high-resolution 3D T2-SPACE imaging, performance further improved, yielding a mean accuracy of 0.80 ± 0.12, sensitivity of 0.67 ± 0.23, specificity of 0.95 ± 0.12, and ROC-AUC of 0.85 ± 0.11. High-resolution T2-SPACE images provided an improvement in specificity compared with conventional T2WI.Conclusions: Radiomic texture features extracted from conventional and high-resolution MRI, when combined with ensemble machine learning methods, show potential for facilitating discrimination between patients with hemifacial spasm and healthy controls. The particularly high specificity achieved with 3D T2-SPACE imaging suggests that this approach may serve as a promising complementary diagnostic tool, supporting conventional imaging and clinical assessment and offering a step toward objective, image-based biomarkers for HFS.

Background: Impaired balance is one of the most common and functionally limiting problems in children with cerebral palsy (CP), significantly affecting their motor abilities and quality of life. Although force platforms are considered the gold standard for evaluating postural stability, they are often costly, non-portable, and require specialized laboratory environments, limiting their accessibility in routine clinical settings. Objective: This study aimed to develop a novel software program based on image processing techniques to assess static balance in children with CP and to evaluate its validity against traditional force platform measurements. Methods: A total of 83 children aged 5–15 years (63 with CP, GMFCS levels I–II; 20 healthy controls) participated. Static balance was assessed under four different standing conditions using both a force platform and a newly developed video-based software tool. The software utilized the frame difference method to detect center of mass movements, and parameters such as velocity and total displacement were calculated. Correlation analyses were conducted between the image processing and force platform data. Results: The software demonstrated moderate to strong positive correlations with force platform parameters in the majority of test conditions, particularly when participants stood with eyes open. In more challenging balance scenarios (e.g., eyes closed, feet together), correlations were weaker but still significant. Conclusions: The findings suggest that this image-based software is a valid, low-cost, and portable alternative for static balance assessment in children with CP. It has the potential for use in diverse clinical or home settings, supporting individualized rehabilitation strategies.

ÖZET AMAÇ: Gelişimsel venöz anomali (GVA), en yaygın serebral vasküler malformasyon türüdür ve genellikle tesadüfen teşhis edilir. Manyetik rezonans görüntüleme (MRG) teknolojisinin özel bir uygulaması olan duyarlılık ağırlıklı görüntüleme (SWI); mikrokanamalar, demir birikimleri ve iskemik lezyonların tespitinde yaygın olarak kullanılır. Bu çalışmada, derin öğrenme tabanlı bir modelin GVA tanısındaki doğruluğunu belirlemek ve klinik uygulamalardaki potansiyel kullanımını değerlendirmek amaçlanmıştır. GEREÇ VE YÖNTEM: Bu çalışmaya, Ocak 2021 - Mayıs 2023 tarihleri arasında hastanemizde kraniyal SWI MRG incelemesi yapılan ve GVA saptanan 99 hasta ile GVA bulunmayan 100 kontrol dahil edilmiştir. Tüm görüntüleme verileri, 10 yıllık deneyime sahip bir nöroradyolog tarafından değerlendirilmiştir. Derin öğrenme süreci, DenseNet121 modeli kullanılarak başlatılmıştır. BULGULAR: Çalışma kohortu 109 kadın ve 90 erkekten oluşmakta olup, ortalama yaş 41,62 ± 19,69 yıl olarak hesaplanmıştır. GVA tanısı alan 99 hastada toplam 104 lezyon tespit edilmiştir. Geliştirilen modelin GVA'ları tespit etmede duyarlılığı (%85 ± 5,0), özgüllüğü (%81 ± 9,6), doğruluğu (%83 ± 5,7) ve eğri altındaki alanı (%90 ± 2,99) yüksek bulunmuştur. TARTIŞMA: Bu çalışmanın bulguları, geliştirilen derin öğrenme modelinin GVA'ların doğru tanısında etkili bir şekilde kullanılabileceğini göstermektedir. Model, literatürde önemli bir boşluğu doldurmakta ve gelecekteki araştırmalar için sağlam bir temel sunmaktadır. Farklı popülasyonlarda test edilmesi ve harici doğrulama çalışmalarının yapılması, modelin genellenebilirliğini ve güvenilirliğini daha da artıracaktır. ABSTRACT AIM: Developmental venous anomalies (DVAs) represent the most common type of cerebral vascular malformation and are typically diagnosed incidentally. Susceptibility-weighted imaging (SWI), a specialized application of magnetic resonance imaging (MRI) technology, is extensively utilized to detect microhemorrhages, iron deposits, and ischemic lesions. This study aimed to determine the accuracy of a deep learning-based model for diagnosing DVAs and to assess its potential applicability in clinical practice. MATERIAL AND METHOD: This study included 99 patients with DVAs detected on cranial SWI MRI conducted at our hospital between January 2021 and May 2023, as well as 100 controls without DVAs. All imaging data were evaluated by a neuroradiologist with 10 years of experience. The deep learning process was initiated using the DenseNet121 model. RESULTS: The study cohort consisted of 109 women and 90 men, with a mean age of 41.62 ± 19.69 years. A total of 104 lesions were identified in 99 patients diagnosed with DVAs. The developed model demonstrated high sensitivity (85% ± 5.0), specificity (81% ± 9.6), accuracy (83% ± 5.7), and area under the curve (90% ± 2.99) values in detecting DVAs. CONCLUSION: The findings of this study indicate that the developed deep learning model can be effectively utilized for the accurate diagnosis of DVAs. This model addresses a significant gap in the literature and provides a robust foundation for future research. Testing the model across different populations and conducting external validation studies will further enhance its generalizability and reliability.

Summary: Extreme learning machine (ELM) is a prominent example of neural network with its fast training speed, and good prediction performance. An online version of ELM called online sequential extreme learning machine (OS‐ELM) has also been proposed for the sequential training. Combined with the need for regularization to prevent over‐fitting in addition to the large number of neurons required in the hidden layer, OS‐ELM demands huge amount of computation power for the large‐scale data. In this article, a mixed norm () regularized online machine learning algorithm (MRO‐ELM) that is based on alternating direction method of multipliers (ADMM) is proposed. A linear combination of the mixed norm and the Frobenius norm regularization is applied using the ADMM framework and update formulas are derived. Graphics processing unit (GPU) accelerated version of MRO‐ELM (GPU‐MRO‐ELM) is also proposed to reduce the training time by processing appropriate parts in parallel using the implemented custom kernels. In addition, a novel automatic hyper‐parameter tuning method is incorporated to GPU‐MRO‐ELM using progressive validation with GPU acceleration. The experimental results show that the MRO‐ELM algorithm and its GPU version outperform OS‐ELM in terms of training speed, and testing accuracy. Also, compared to the cross validation, the proposed automatic hyper‐parameter tuning demonstrates dramatical reduction in the tuning time.