Lightweight radiomics tool spots stroke types on CT scans in pilot study

The framework, described as a proof of concept, combined 16 texture-based measurements from non-contrast CT images with a Random Forest classifier and achieved a validated mean accuracy of 0.9928 (± 0.0101) under patient-level stratified Group K-fold cross-validation. Precision reached 1.00, while recall and F1-scores ranged between 0.99 and 1.00 across all classes within the validation setup. The team, led by Etienne Jiazet and colleagues, analyzed an anonymized dataset from the Bafoussam Regional Hospital Center and published the work in the Bulletin of the National Research Centre. The authors are careful to frame the results as a preliminary demonstration of feasibility rather than a validated diagnostic system, but the numbers and the modest computing requirements behind them have drawn attention to a persistent gap in medical imaging: tools that work well in well-funded hospitals are often impractical in places with limited infrastructure.

Stroke remains one of the most demanding emergencies in medicine. The study’s authors note that stroke ranks as the third leading cause of death worldwide and generates substantial annual economic costs for post-stroke patient management. Among imaging options, non-contrast computed tomography is preferred for initial diagnosis due to its speed, availability, and cost-effectiveness compared with MRI and perfusion CT. Yet manual interpretation of CT images can be time-consuming and subject to inter-observer variability, particularly in resource-limited clinical settings with high workloads. Many existing automated analysis tools rely on complex architectures that are challenging to deploy in such environments, while advanced deep learning models function as computational black boxes and demand heavy hardware infrastructure, such as high-end GPUs, that resource-constrained settings simply do not have.

The new framework was designed specifically to occupy the middle ground: lightweight enough to run on ordinary hardware, yet discriminative enough to separate stroke subtypes. The dataset came from brain CT scans acquired at the Bafoussam Regional Hospital Center using a Siemens Healthineers SOMATOM go.Now scanner with 16 detector rows. The analyzed slices were reconstructed at a slice thickness of 0.80 mm, with a standard matrix resolution of 512 × 512 pixels, a field of view ranging from 220 to 240 mm, and a raw dynamic range of −1024 to +3071 Hounsfield Units before windowing. The dataset included 66 patients, each contributing four to six CT slices selected by the supervising radiologist based on clinical relevance. The lesions primarily involved the middle cerebral artery territory, the basal ganglia, and the main cerebral lobes, and the selection criteria focused on macroscopically visible, clinically significant lesions while excluding micro-lesions. All scans were performed as emergency-grade imaging immediately upon patient admission in the acute phase of the stroke.

In total, 330 images were included, categorized into ischemic stroke, hemorrhagic stroke, or normal cases with no detectable brain lesions. To prevent data leakage, the dataset was partitioned at the patient level rather than the image level, using a stratified Group K-Fold cross-validation scheme with K = 5. For each fold, four partitions served for training and one for testing, ensuring that all CT slices from a given patient remained exclusively within the same set—an important safeguard, since splitting slices from the same patient across training and test sets can inflate performance estimates. The machine learning pipeline was implemented in Python and executed on a desktop computer equipped with an Intel Core i5 processor running at 2.67 GHz and 8 GB of RAM, hardware far below what deep learning pipelines typically require. All patient-identifying data were anonymized, though the DICOM images retained acquisition parameters such as window center and window width, which are essential for accurate rendering and reliable texture analysis.

Preprocessing played a deliberate role in sharpening the signals the model would later exploit. Each DICOM image underwent clinical windowing using a window width of 80 HU and a window level of 35 HU, isolating the [−5, +75] HU range that targets the standard soft-tissue brain window. The images were then normalized to 8-bit grayscale to ensure consistency in the intensity values used for subsequent analyses. According to the authors, this windowing operation does not destroy texture information but optimizes its statistical boundaries: in unwindowed data, subtle soft-tissue variations are heavily compressed by the extreme densities of the skull and air, whereas restricting the dynamic range stretches the density profiles of the tissues and renders their statistical characteristics more discriminative. The histograms reflected the underlying pathology. Hemorrhagic tissue exhibited a distinct rightward shift capturing the hyperdensity of acute blood, ischemic tissue shifted markedly to the left, highlighting the hypodensity caused by cytotoxic edema, and normal brain parenchyma maintained a stable, symmetric distribution centered around 30 to 35 HU.

A distinctive feature of the workflow is its reliance on manually targeted regions of interest. Circular regions with a fixed radius of 12 pixels—corresponding to approximately 10.8 mm in physical diameter—were positioned by expert radiologists using a single-click interactive interface on the windowed CT images. Because the framework performs feature extraction and classification strictly in two dimensions on a slice-by-slice basis, each ROI is processed independently, making the analysis invariant to inter-slice thickness or potential slice gaps. The restricted size was chosen to capture a representative homogeneous portion of tissue while avoiding boundary effects from surrounding bone, cerebrospinal fluid, and peri-lesional edema. From each ROI, the team extracted 16 raw textural features using the PyRadiomics and Scikit-Image libraries, adhering to the mathematical standards of the Image Biomarker Standardisation Initiative. The features fell into two complementary categories: first-order statistical descriptors, which capture global intensity distributions, and second-order textural descriptors derived from Gray-Level Co-occurrence Matrices, which capture local spatial relationships between pixels.

The quantitative analysis revealed distinct digital signatures for each class. The hemorrhagic class showed higher mean intensity (52.275) and maximum intensity (104.135) than the ischemic (mean 10.268) and normal (mean 26.997) classes, consistent with the bright appearance of acute blood on CT, while the ischemic class presented the lowest values, reflecting the hypodense appearance of infarcted tissue and edema. Hemorrhagic lesions also exhibited the highest GLCM contrast (2658.842), reflecting abrupt transitions between hyperdense blood and surrounding parenchyma, whereas ischemic regions displayed the highest first-order energy and the lowest entropy, markers of textural uniformity. Feature importance analysis using Random Forest Gini importance showed that first-order statistics dominated: standard deviation (21.09%), median (20.80%), maximum intensity (17.29%), and mean (16.08%) together accounted for 75.26% of the model’s weight, with GLCM features such as dissimilarity (4.27%) and contrast (3.52%) contributing secondary decision support for overlapping boundary cases.

The classification results were strong within the validation setup. The Random Forest ensemble, configured with 100 trees and a square-root feature subset at each split, achieved a validated average accuracy of 99.28% with a 95% confidence interval of [98.03%, 100.00%], average precision of 1.00, and an AUC close to 1.00. In a representative cross-validation fold of 66 test images, exactly two discordant cases appeared: one ischemic case misclassified as normal and one normal case misclassified as ischemic, with no confusion between the hemorrhagic and ischemic classes. An ablation study showed that first-order features alone achieved an accuracy of 0.9879 ± 0.011 and GLCM features alone 0.9424 ± 0.039, while the combined set performed best. Against baseline classifiers including support vector machines, k-nearest neighbors, and decision trees, the Random Forest model achieved the highest average accuracy, though the differences compared with high-performing baselines did not reach statistical significance (p > 0.05) due to high baseline scores.

The authors are explicit about what these results do and do not establish. The system is a targeted decision-support tool focused strictly on ROI classification rather than end-to-end automated detection, and the manual selection of regions of interest, although performed by experts, poses a challenge for full automation and reproducibility. Formal robustness analyses regarding ROI spatial variability and inter-observer variability were not conducted, meaning minor operator-dependent positioning shifts could influence feature stability. The relatively modest sample size from a single center using a single CT scanner model under a standardized 0.80 mm reconstruction protocol limits immediate generalizability, and the lack of external validation on an independent dataset remains a key limitation. Future work will focus on validating the findings on heterogeneous, multi-vendor datasets from other hospital centers, integrating automatic segmentation to bypass manual target selection, and assessing generalizability for real-world clinical translation. As a proof of concept, the study demonstrates that carefully chosen texture descriptors and a modest classifier can perform impressively on a single-center cohort—and that the path from such internal validation to clinical deployment still requires the rigorous external testing the authors themselves call for.

Subject of Research: A lightweight radiomics framework for classifying ischemic stroke, hemorrhagic stroke, and normal findings on non-contrast brain CT scans

Article Title: A lightweight texture-based framework for stroke classification: a radiomics approach in resource-limited clinical settings

Article References: Jiazet, E., Feudjio, G., Tsafack, E. A., Ikamaise, V., & Egbe, N. O. (2026). A lightweight texture-based framework for stroke classification: a radiomics approach in resource-limited clinical settings. Bulletin of the National Research Centre, 50(1), Article 111. https://doi.org/10.1186/s42269-026-01501-z

Image Credits: AI Generated

DOI: 10.1186/s42269-026-01501-z

Keywords: radiomics, stroke classification, non-contrast CT, texture analysis, Random Forest, GLCM, decision support, resource-limited settings, cross-validation, proof of concept, medical imaging, Cameroon