Research on stroke detection algorithm based on the improved model of YOLOv8-CBAM-ASFF

Accurate early diagnosis of stroke is critical for reducing mortality and disability. Leveraging the rapid advances in deep learning, automated stroke detection has become an active research focus. Yet, existing approaches still underperform in detecting small lesions, adapting to complex backgrounds, and operating in real time. Four YOLOv8 variants of different scales were systematically evaluated, and an optimized architecture integrating the Convolutional Block Attention Module (CBAM) and the Adaptive Spatial Feature Fusion (Detect_ASFF) component was proposed. CBAM selectively amplifies salient features while suppressing background noise, whereas Detect_ASFF dynamically reweights multi-scale features, mitigating small-object loss and contextual degradation. Compared with the baseline, the refined model achieved a comparable overall mAP while improving lesion-level sensitivity for medium and large lesions. On a strictly patient-level held-out test set (34 patients, 474 images), the enhanced YOLOv8n reached 74.9% mean average precision at an intersection-over-union threshold of 0.5, where a lesion counts as detected when the predicted box overlaps the annotated lesion by at least 50% (mAP@0.5), and 36.1% when the same measure is averaged over thresholds from 0.5 to 0.95 (mAP@0.5:0.95). It also increased the recall of medium-sized lesions (32–96 px) from 62.5% to 73.5% and of large lesions from 84.9% to 90.3%, at a modest cost of 0.2 additional false positives per image. Owing to its lightweight design, the optimized YOLOv8n is a promising candidate for real-time stroke-lesion detection; this work is presented as a proof of principle, and clinical deployment requires further validation.