Attention-enhanced CNNs classify four stroke types on two-center brain CT

Squeeze-and-excitation attention added to DenseNet-121, ResNet-50, and EfficientNet-B3 improved four-class stroke classification on 5000 two-center cranial CTs by an average 0.93 percentage points, with largest gains separating ischemic from hemorrhagic and acute from chronic is…

• Dataset: 5000 cranial CT images from two centers, reviewed by radiology specialists and labeled as normal, hemorrhagic stroke, acute ischemic stroke, or chronic ischemic stroke.
• Mean classification accuracy gain was 0.93 percentage points across all three SE-enhanced architectures versus standard counterparts.
• Improvement was most pronounced for hemorrhagic-versus-ischemic and acute-versus-chronic ischemic distinctions.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.