Title : Interpretable domain-specific deep learning for multiclass stroke classification using CT scan images

Stroke classification using CT scan images is critical for enabling timely diagnosis and effective clinical decision-making. However, automated detection of stroke types—particularly early-stage ischemic stroke—remains challenging due to subtle variations in tissue density that are often difficult to capture. Most existing deep learning approaches rely on models pre-trained on ImageNet, a dataset of natural images optimized for object recognition, which limits their effectiveness in medical imaging tasks.

In this study, we propose an interpretable, domain-specific deep learning framework for multiclass stroke classification, including ischemic, hemorrhagic, and normal cases. The framework utilizes a fixed feature extractor based on a Medical Masked Autoencoder (MedMAE), pre-trained on large-scale medical imaging data, to learn high-level and domain-relevant feature representations. To further enhance performance, we introduce an adaptive skull stripping method that removes non-brain structures and reduces imaging noise, allowing the model to focus on clinically relevant regions. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) is employed to generate visual attention maps, improving model interpretability and transparency.

The proposed model was evaluated on a dataset of 6,653 CT scan slices. Experimental results demonstrate that the MedMAE-based approach achieved an accuracy of 92.1%, outperforming the ImageNet-based baseline model, which achieved 86.7%. Notably, the proposed method significantly improved ischemia detection, with the F1 score increasing from 0.63 to 0.88. Furthermore, qualitative analysis revealed that the baseline model often relied on irrelevant features near the skull boundary, whereas the proposed framework focused more accurately on stroke pathology regions.

These findings highlight the importance of domain-specific pretraining and interpretability in medical imaging, offering a more reliable and clinically meaningful approach for automated stroke classification.

Bashemera Brenda Birungi is a PhD student specializing in Artificial Intelligence for Health. Her research focuses on applying deep learning techniques to medical imaging for improved diagnosis and clinical decision support, particularly in stroke detection using CT scans. She has experience in developing interpretable AI models and is passionate about bridging the gap between machine learning and real-world healthcare applications, especially in resource-constrained settings.