Early Diagnosis of Ischemic Stroke on Non-Contrast CT Scans Using a Convolutional Neural  Network: A Case Study from Mulago National Referral Hospital, Uganda

Technical Summary: Early Diagnosis of Ischemic Stroke on Non-Contrast CT Scans Using a CNN in Uganda

Problem Statement
In low- and middle-income countries (LMICs), particularly Uganda, the diagnosis of ischemic stroke is frequently delayed due to a scarcity of specialist radiologists. While non-contrast computed tomography (NCCT) is the most widely available imaging modality for suspected stroke, detecting early ischemic changes on these scans is challenging and typically requires experienced interpretation. In public hospitals like Mulago National Referral Hospital, reporting delays of three to five days are common as radiologists prioritize trauma and other emergencies. Existing computer-assisted diagnostic solutions often rely on advanced imaging (e.g., CT perfusion or angiography) unavailable in resource-limited settings or fail to provide point-of-care interfaces usable by non-technical staff.

Methodology
This study employed a retrospective diagnostic accuracy design to develop and validate a Convolutional Neural Network (CNN) for early ischemic stroke detection using exclusively NCCT scans from Mulago National Referral Hospital.

• Dataset: The study utilized 1,000 anonymized NCCT scans (500 confirmed ischemic stroke, 500 controls) collected between January 2022 and June 2023. Images were curated from a total pool of 1,450 scans after excluding those with motion artifacts, incomplete series, or corrupted metadata.
• Annotation: Three radiologists with over five years of neuroimaging experience independently annotated the 500 stroke-positive scans, targeting early signs such as hypoattenuation, sulcal effacement, and loss of gray–white matter differentiation. Inter-rater reliability was strong (Cohen's κ = 0.83), with discrepancies resolved by consensus to create the reference standard.
• Preprocessing: The pipeline included intensity normalization, resizing to standardized dimensions, and data augmentation (rotations, translations, flips). Quantitative features extracted included intensity metrics, shape descriptors, and texture features derived from Gray Level Co-occurrence Matrix (GLCM) properties.
• Model Architecture: A 2D CNN was developed in PyTorch to analyze individual slices. The architecture consisted of convolution and pooling layers for feature extraction, followed by fully connected layers and a softmax output. Techniques to prevent overfitting included dropout, L2 regularization, and batch normalization. The model was trained using the Adam optimizer over 100 epochs with GPU acceleration.
• Evaluation Strategy: Performance was assessed using five-fold cross-validation (80% training, 20% testing). Additionally, a comparative analysis was conducted on a strictly held-out independent subset of 30 cases (15 stroke, 15 non-stroke) against a senior consultant radiologist.
• Deployment: A web-based interface was developed using a RESTful API (Flask backend, HTML/CSS/JavaScript frontend) to allow secure upload of DICOM, JPEG, or PNG files and display automated inference results.

Key Results

• Model Performance: Across five-fold cross-validation, the CNN achieved a mean accuracy of 90.2% (±1.8), sensitivity of 91.4% (±1.7), specificity of 89.0% (±2.1), and an Area Under the Curve (AUC) of 91.3% (±1.7). Performance remained stable across validation folds.
• Radiologist Comparison: On the independent 30-case subset, the senior radiologist achieved 100% sensitivity and 95% specificity, missing zero cases. The CNN achieved 86.7% sensitivity and 88.7% specificity, missing two stroke cases. Statistical analysis (McNemar's test) indicated that the difference in diagnostic performance was not statistically significant (p = 0.21).
• Efficiency: The most significant differentiator was processing time. The CNN processed scans in approximately 3 seconds per case, compared to the 5 minutes required for manual radiologist interpretation.
• Platform Functionality: The web-based interface successfully demonstrated the ability to upload scans, perform automated analysis, and generate results within seconds in a simulated hospital environment.

Key Contributions

1. Locally Trained Model: The study presents a CNN trained exclusively on data from a Ugandan national referral hospital, addressing the specific imaging characteristics and resource constraints of the region.
2. Reliance on Routine Imaging: Unlike many advanced AI stroke models, this system operates solely on non-contrast CT scans, the most accessible modality in LMICs, rather than requiring perfusion or angiography.
3. Integrated Deployment: The research moves beyond experimental algorithms by delivering a functional, browser-based platform with a RESTful API, designed for use by non-technical healthcare staff.
4. Validation in Resource-Limited Context: The study validates that high diagnostic performance is achievable in settings with limited radiology expertise, offering a potential tool for rapid triage.

Significance and Claims
The authors claim that the integrated system offers diagnostic accuracy comparable to that of a senior radiologist while delivering results significantly faster. The study suggests that such a platform has promising potential to support stroke triage and reduce diagnostic delays in settings where specialist radiologists are scarce. The authors emphasize that the system addresses the critical bottleneck of delayed diagnosis in emergency care.

However, the paper maintains a modest tone regarding its current impact. The authors explicitly state that the platform has not yet undergone live clinical deployment and was tested only in a simulated environment with anonymized data. They acknowledge limitations including the single-center nature of the data, the lack of visual interpretability features (e.g., heatmaps), and the absence of structured feedback from end-users. Consequently, the authors conclude that further multicenter validation and real-world pilot implementation are necessary before large-scale deployment can be recommended.