Diferentiating multiple sclerosis (MS) from ischemic stroke lesions on MRI remains a clinical challenge due to their similar appearances as white matter hyperintensities. We propose a radiomics-based machine learning framework that integrates multi-level explainable AI (XAI) techniques to support transparent and clinically meaningful lesion classification. Radiomic features are extracted from standardized MRI scans and used to train multiple classifiers, with Random Forest achieving the best performance (accuracy: 91.24%, F1: 86.54%). The framework incorporates four complementary explanation layers: global insights using SHAP, local interpretability via LIME, counterfactual reasoning with DiCE, and clinical narrative generation using GPT-based language models. This layered approach enhances interpretability at both dataset and lesion levels, enabling clinicians to understand, trust, and act upon model outputs. A radiologist who reviewed the results found the explanations helpful and confirmed that the overall analysis was clinically meaningful. Our results demonstrate the value of combining radiomics and advanced XAI techniques for diferential diagnosis of brain lesions.
Diferentiating multiple sclerosis (MS) from ischemic stroke lesions on magnetic resonance imaging
(MRI) is a complex diagnostic task due to their overlapping appearance as white matter hyperintensities
(WMHs). MS is a chronic inflammatory disease marked by demyelination, while ischemic lesions arise
from vascular occlusion and subsequent tissue damage. Despite their distinct pathologies, both appear
similar on common MRI sequences like FLAIR, complicating manual diagnosis and often requiring
expert interpretation [
Radiomics ofers a quantitative approach to analyze lesion characteristics by extracting texture,
shape, and intensity-based features from medical images [
Explainable AI (XAI) techniques have developed to tackle this issue by making model decisions
interpretable. SHapley Additive exPlanations (SHAP) [
In contrast to previous work, our study proposes a multi-level XAI framework that integrates four layers of interpretability: (1) SHAP for global feature importance, (2) LIME for local explanations, (3) DiCE for counterfactual reasoning, and (4) GPT-generated clinical narratives for human-aligned interpretation. This layered approach supports both technical transparency and clinical relevance. We evaluate the method using lesion-wise radiomic features extracted from two public datasets—MSSEG (for MS) and ISLES (for stroke)—and compare multiple classifiers, identifying Random Forest as the best-performing model. To the best of our knowledge, this is the first radiomic framework to combine SHAP, LIME, counterfactual explanations, and language-based narratives for transparent brain lesion classification.
This study utilizes two publicly available MRI datasets for lesion segmentation and classification. The
ifrst is the ISLES 2022 dataset [
All MRI volumes were converted into 2D slices along axial, coronal, and sagittal planes. The following
preprocessing steps were applied to each slice:
• Bias Field Correction: N4 bias correction was applied to reduce intensity non-uniformities.
• Intensity Normalization: Pixel intensities were scaled to the [0, 255] range using min–max
normalization.
• Denoising: Non-local means filtering [
Radiomic features were extracted on a per-lesion basis using the PyRadiomics library [
We evaluated three classifiers: Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM). Each model was optimized using RandomizedSearchCV (method from scikit-learn python library) on a combined training and validation set with a predefined split. Hyperparameters were selected based on F1-score. The final models were retrained on the full training+validation set and evaluated on the held-out test set.
To ensure transparency and clinical relevance, the proposed framework integrates four complementary
XAI strategies:
• Global Explanations (SHAP): SHAP values [
This layered approach supports both technical and human-aligned interpretability, enhancing transparency and trust in the AI-assisted diagnosis process.
We evaluated three classifiers on the test set and analyzed the interpretability of the best-performing model using SHAP, LIME, DiCE, and GPT-based narratives.
SHAP values were computed to explain the contribution of each radiomic feature across the dataset. Figure 1 shows that texture features like glcm_Idmn and firstorder_Skewness were consistently important.
LIME provided per-lesion explanations, and GPT transformed these into clinical narratives. Figure 2 illustrates a case predicted as MS. High uniformity and low contrast were influential, supporting the MS diagnosis. In contrast, Figure 3 shows a predicted ischemic lesion. High skewness and entropy supported the prediction, indicating structural heterogeneity.
To simulate alternative diagnostic scenarios and examine model robustness, we used DiCE to generate counterfactual examples. These identify minimal, actionable changes to radiomic features that would result in a diferent prediction outcome. This approach enhances transparency by answering the question: "What would need to change for the lesion to be classified diferently?"
Figure 4 shows a counterfactual explanation for a lesion originally classified as MS. DiCE suggests that reducing the mean intensity (original_firstorder_Mean) from 97.45 to 31.60, and increasing the Predicted Diagnosis: MS (Multiple Sclerosis) – Lesion 3 Prediction Probability: 97% Key Features Supporting MS: •High Gray Level Run Emphasis (1.70): Indicates concentrated areas with high signal intensity runs, commonly seen in MS plaques due to demyelination and chronic inflammation. •High Gray Level Zone Emphasis (2.50): Reflects large, homogeneously intense regions, consistent with typical MS lesion appearance in FLAIR MRI. •Gray Level Variance (0.18): Low heterogeneity in gray levels suggests uniform lesion intensity, a hallmark of MS plaques. •Low Gray Level Zone Emphasis (0.62): Reflects reduced presence of darker, ischemia-associated zones, favoring MS. •Short Run Low Gray Level Emphasis (0.36): Indicates less fragmented low-signal areas, again pointing toward uniform MS lesions. •First-order Median Intensity (116.00): Falls within the expected range for MS lesions, indicating moderately hyperintense signal. Why Ischemic Stroke is Unlikely: •Autocorrelation (1.67): Below typical ischemic levels; ischemic lesions often show higher internal texture correlation. •Joint Average (1.27) & Sum Average (2.55): These GLCM features are lower than typical ischemic values, indicating less internal gray-level co-occurrence and spread. •NGTDM Complexity (0.25): Much lower than expected in ischemic strokes, which usually present with more complex and heterogeneous textures.
Conclusion: Lesion 3 is classified as MS primarily due to its homogeneous intensity, lack of low-gray regions, and texture simplicity, which are characteristic of chronic demyelinating lesions rather than ischemic damage. original_glszm_SmallAreaHighGrayLevelEmphasis from 0.14 to 33.46, would be suficient to lfip the prediction to ischemic. These feature adjustments reflect plausible variations in lesion brightness and structural homogeneity that align with known imaging patterns of ischemic pathology. This layer of "what-if " analysis enables clinicians to explore how small radiomic shifts could afect classification outcomes, making the model more interpretable and clinically interactive. Although DiCE successfully generated plausible counterfactual examples, interpreting the proposed feature modifications requires clinical expertise, as some radiomic changes may not correspond directly to observable anatomical changes.
Predicted Diagnosis: Ischemic Stroke – Lesion 1 Prediction Probability: 87% Key Features Supporting Ischemic Stroke: •High Gray Level Zone Emphasis (7.80): Indicates large, bright uniform regions—typical of acute/subacute infarcts with welldemarcated necrotic tissue. •Short Run Low Gray Level Emphasis (0.12): Suggests a lack of short, low-intensity textures—supporting ischemia, where lesions tend to be sharply defined and not mottled. •Low Gray Level Zone Emphasis (0.31): Minimal presence of dark areas aligns with the uniform brightness of ischemic lesions. •First-order Median (163.00), Root Mean Squared (170.14), and Mean Intensity (169.25): All indicate high signal intensity, characteristic of ischemic tissue with edema or gliosis in FLAIR MRI.
Why MS is Unlikely: •Joint Average (2.43): While elevated, it's less distinctive for MS compared to more specific texture patterns. •Autocorrelation (6.15): Although relatively high, it does not offset the dominant ischemic characteristics. •NGTDM Complexity (1.55): Moderate complexity, but not high enough to suggest the varied texture seen in MS plaques. •Gray Level Variance (1.04): Suggests some heterogeneity but insufficient to match the typical diverse texture of MS lesions. Conclusion: Lesion 1 is classified as Ischemic Stroke due to its high uniform brightness, low texture fragmentation, and elevated mean intensities, all of which are classic features of infarcts rather than demyelinating MS lesions. The combination of radiomics and multi-level XAI revealed consistent and interpretable patterns in lesion classification. SHAP highlighted globally dominant features, while LIME showed per-lesion factors contributing to predictions. DiCE further provided hypothetical scenarios for counter-diagnosis, and GPT-based narratives completed the pipeline by ofering concise, clinician-aligned explanations. Together, these methods support not only technical validation but also clinical usability, making the system suitable for diagnostic support in real-world settings. Each explanation modality brings distinct strengths and limitations. SHAP provides both global and local feature attributions but can be computationally intensive. LIME generates interpretable local explanations but may sufer from instability across small perturbations. DiCE enables actionable counterfactuals but sometimes proposes feature changes that may not correspond directly to plausible anatomical variations. GPT narratives ofer intuitive, clinician-friendly summaries, but they are prone to occasional hallucinations, especially in ambiguous cases. Understanding these trade-ofs is crucial for practical clinical adoption.
To assess the clinical relevance of the generated explanations, we shared the model outputs and narratives with a practicing radiologist. The radiologist noted that the explainability provided was very helpful and that the overall analysis made sense. However, broader evaluation is needed. Future work will involve multi-expert validation with inter-rater agreement scoring and quantitative assessment of explanation fidelity to strengthen the clinical robustness of the framework. Additionally, integrating counterfactual outputs from DiCE into GPT-based narrative generation presents an exciting opportunity to create "what-if" clinical explanations that can further enhance the usability of the system.
This study presents a radiomics-based classification framework for diferentiating multiple sclerosis and ischemic stroke lesions on MRI, enhanced with a multi-level explainable AI pipeline. Among the evaluated classifiers, Random Forest achieved the best overall performance, with an accuracy of 91.2% and an F1-score of 86.5%. Beyond predictive performance, the framework integrates global (SHAP), local (LIME), counterfactual (DiCE), and natural language (GPT) explanations to provide transparent and clinically meaningful insights. SHAP identified texture-based radiomic features as globally influential, while LIME and GPT enabled per-lesion interpretability in clinician-friendly language. DiCE ofered hypothetical reasoning to explore how small changes in feature values could lead to diferent diagnoses. Together, these methods create a robust and interpretable decision support system that bridges technical and clinical domains. Future work will focus on experimenting with diferent radiomic extraction parameters, such as varying angles, distances, and bin widths, to explore their impact on classification and explainability. We also plan to expand this framework to include multi-modal MRI inputs, increase sample diversity for better generalization, and integrate visual explanation methods, such as attentionbased heatmaps or Grad-CAM visualizations, to enhance clinical trustworthiness. Finally, fine-tuning the narrative generation process using domain-specific language models will be explored. By combining radiomics with layered explainability, this approach supports the safer, more transparent deployment of AI tools in neuroimaging diagnostics.
This research is funded by the SFI Center for Research Training in Machine Learning (18/CRT/6183).
During the preparation of this work, the author(s) used GPT-4o and Grammarly for grammar and spelling checks. The author(s) reviewed and edited the output as necessary and assume full responsibility for the content of the publication.