Bio-Inspired Dung Beetle Optimization-Driven InceptionNeXt-Tiny–Swin Transformer Network with Adaptive Dual Attention for Brain Hemorrhage Classification Using CT Images

Intracranial hemorrhage (ICH) signifies bleeding occurring within the brain tissue or surrounding intracranial spaces, necessitating timely and accurate diagnosis to reduce the risk of death and lasting neurological deficits. Although computed tomography (CT) imaging is broadly used for hemorrhage detection, manual evaluation of large numbers of CT scans is tedious and susceptible to variability in clinical interpretation. To overcome these challenges, this research proposes a bio-inspired Dung Beetle Optimization (DBO)-driven InceptionNeXt-Tiny–Swin Transformer network with Adaptive Dual Attention (ADA) for automated ICH classification. Initially, CT images undergo preprocessing and augmentation to improve data quality and model generalization. InceptionNeXt-Tiny is employed to extract discriminative multi-scale spatial features, while the Swin Transformer performs hierarchical contextual representation learning and captures longer-range spatial dependencies across neighboring windows. Subsequently, the ADA module refines the learned representations through channel and spatial attention mechanisms, emphasizing hemorrhage-relevant features. Furthermore, DBO is employed to optimize critical hyperparameters, thereby enhancing the convergence and classification performance of the proposed model. The proposed framework was evaluated using the Intracranial Brain Hemorrhage CT Images dataset containing normal and hemorrhagic brain CT scans from 82 patients, which was partitioned at the patient level to prevent information leakage, with patient-level predictions obtained by majority voting over slices. Experimental outcomes on an independent test subset of 18 unseen patients indicate that the model achieves 94.44% accuracy, 100.00% recall, 87.50% precision, 93.33% F1-score, and an AUC of 0.9870. Ablation studies, patient-level cross-validation analysis, and comparisons with conventional models were conducted to evaluate the effectiveness and consistency of the proposed approach, with cross-validation yielding a mean accuracy of 92.65 ± 2.88%. In an external evaluation on 442 scans from the multi-centre CQ500 dataset without retraining, the framework achieved 92.08% accuracy, 92.59% recall, 89.29% precision, and 90.91% F1-score. The results demonstrate that the integration of multi-scale feature extraction, transformer-based contextual representation, adaptive attention refinement, and bio-inspired DBO-based hyperparameter optimization enhances the effectiveness of ICH classification.