Melahat Poyraz

Background/Objectives: The neonatal and infant periods represent a critical window for brain development, characterized by rapid and heterogeneous processes such as myelination and cortical maturation. Accurate assessment of these changes is essential for understanding normative trajectories and detecting early abnormalities. While conventional MRI provides valuable insights, automated classification remains challenging due to overlapping developmental stages and sex-specific variability. Methods: We propose SEPoolConvNeXt, a novel deep learning framework designed for fine-grained classification of neonatal brain development using T1- and T2-weighted MRI sequences. The dataset comprised 29,516 images organized into four subgroups (T1 Male, T1 Female, T2 Male, T2 Female), each stratified into 14 age-based classes (0–10 days to 12 months). The architecture integrates residual connections, grouped convolutions, and channel attention mechanisms, balancing computational efficiency with discriminative power. Model performance was compared with 19 widely used pre-trained CNNs under identical experimental settings. Results: SEPoolConvNeXt consistently achieved test accuracies above 95%, substantially outperforming pre-trained CNN baselines (average ~70.7%). On the T1 Female dataset, early stages achieved near-perfect recognition, with slight declines at 11–12 months due to intra-class variability. The T1 Male dataset reached >98% overall accuracy, with challenges in intermediate months (2–3 and 8–9). The T2 Female dataset yielded accuracies between 99.47% and 100%, including categories with perfect F1-scores, whereas the T2 Male dataset maintained strong but slightly lower performance (>93%), especially in later infancy. Combined evaluations across T1 + T2 Female and T1 Male + Female datasets confirmed robust generalization, with most subgroups exceeding 98–99% accuracy. The results demonstrate that domain-specific architectural design enables superior sensitivity to subtle developmental transitions compared with generic transfer learning approaches. The lightweight nature of SEPoolConvNeXt (~9.4 M parameters) further supports reproducibility and clinical applicability. Conclusions: SEPoolConvNeXt provides a robust, efficient, and biologically aligned framework for neonatal brain maturation assessment. By integrating sex- and age-specific developmental trajectories, the model establishes a strong foundation for AI-assisted neurodevelopmental evaluation and holds promise for clinical translation, particularly in monitoring high-risk groups such as preterm infants.

The main aim of this study is to propose a novel convolutional neural network, named BrainNeXt, for the automated brain disorders detection using magnetic resonance images (MRI) images. Furthermore, we aim to investigate the performance of our proposed network on various medical applications. To achieve high/robust image classification performance, we gathered a new MRI dataset belonging to four classes: (1) Alzheimer's disease, (2) chronic ischemia, (3) multiple sclerosis, and (4) control. Inspired by ConvNeXt, we designed BrainNeXt as a lightweight classification model by incorporating the structural elements of the Swin Transformers Tiny model. By training our model on the collected dataset, a pretrained BrainNeXt model was obtained. Additionally, we have suggested a feature engineering (FE) approach based on the pretrained BrainNeXt, which extracted features from fixed-sized patches. To select the most discriminative/informative features, we employed the neighborhood component analysis selector in the feature selection phase. As the classifier for our patch-based FE approach, we utilized the support vector machine classifier. Our recommended BrainNeXt approach achieved an accuracy of 100% and 91.35% for training and validation. The recommended model obtained the test classification accuracy of 94.21%. To further improve the classification performance, we suggested a patch-based DFE approach, which achieved a test accuracy of 99.73%. The obtained results, surpassing 90% accuracy on the test dataset, demonstrate the effectiveness and high classification performance of the proposed models.