Hybrid ConvNeXt–Vision Mamba with Hemispheric Difference-Aware Attention and Bioinspired Optimization for Stroke Classification

Stroke requires rapid and accurate diagnosis, yet automated interpretation of computed tomography (CT) images remains challenging because of subtle lesion characteristics, anatomical variability, and similarities between normal and pathological tissue patterns. This study presents a hybrid ConvNeXt-Tiny–Vision Mamba framework for binary stroke classification from rendered axial CT slice images, in which ConvNeXt-Tiny extracts hierarchical local features and Vision Mamba captures long-range contextual dependencies. Motivated by the bilateral organization of the brain, an Adaptive Hemispheric Difference Attention (AHDA) module models interhemispheric feature differences through central sagittal partitioning and contralateral reflection, while Adaptive Cross-Attention Fusion (ACAF) integrates the local and asymmetry-enhanced representations. An Adaptive Lévy Flight Gorilla Troops Optimizer (ALGTO), extending Gorilla Troops Optimization with nonlinear parameter adaptation and stagnation-triggered Lévy-flight perturbation, performs hyperparameter optimization. A public Kaggle dataset of 2501 unique CT slice images from 82 filename-defined case groups was used; the images are provided as rendered files without DICOM metadata, and the correspondence of case groups to distinct patients could not be verified. All partitioning was performed at the case-group level. On the locked, single-use internal case-level test set (17 case groups), the framework achieved an accuracy of 94.12%, a weighted F1-score of 94.20%, and an area under the receiver operating characteristic curve (ROC-AUC) of 0.9697, and case-grouped five-fold cross-validation on 65 development case groups yielded a mean accuracy of 90.77 ± 3.44%, with mean precision of 89.33 ± 9.83% and F1-score of 87.92 ± 4.54%. Ablation analyses on the development set showed performance gains with the addition of Vision Mamba and AHDA, more consistent performance across folds with ACAF, and a modest further improvement with ALGTO, and Gradient-weighted Class Activation Mapping (Grad-CAM) visualization provided a qualitative illustration of the image regions influencing predictions. The framework demonstrates the potential of combining anatomically motivated attention with bioinspired adaptive optimization for CT-based stroke classification on this dataset.