StrokeCT-2C5K: A Two-Center Cranial CT Dataset for Four-Class Stroke Classification Using SE-Attention-Enhanced Deep Learning Models

Résumé fourni par la source

Stroke is one of the leading causes of mortality and long-term neurological disability worldwide, and early diagnosis through accurate disease classification directly affects treatment success. Rapid differentiation of hemorrhagic and ischemic stroke on computed tomography (CT) images, together with accurate determination of the acute and chronic phase in ischemic cases, is of critical importance in the clinical decision-making process. In this study, StrokeCT-2C5K a two-center dataset comprising 5000 cranial CT images, was assembled specifically for this work. The images were reviewed by radiology specialists and assigned to one of four diagnostic categories: normal, hemorrhagic stroke, acute ischemic stroke, or chronic ischemic stroke. A Squeeze-and-Excitation (SE-Attention) mechanism was then integrated into DenseNet-121, ResNet-50, and EfficientNet-B3. From a biomimetic perspective, this channel-recalibration process provides a functional analogy to biological selective attention by giving greater weight to informative responses while reducing the influence of less relevant ones. All models were trained under the same training, validation, and test protocol; the standard CNN architectures were compared with their SE-Attention-enhanced counterparts. The results showed that the SE-Attention mechanism enables more effective learning of lesion-specific discriminative features by adaptively recalibrating channel-wise information, yielding an average classification accuracy improvement of 0.93 percentage points across all three architectures. The most pronounced improvements were observed in distinguishing ischemic from hemorrhagic stroke, as well as in distinguishing acute from chronic ischemic stroke. These findings show that a selective-information-processing strategy functionally analogous to biological attention can improve multi-class stroke classification across different CNN backbones.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.