Comparing Baseline and Day-1 Diffusion MRI Using Multimodal Deep Embeddings for Stroke Outcome Prediction

arXiv:2512.02088 · eess.IV, cs.AI, cs.CV, cs.LG · Submitted 2025-12-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Comparing Baseline and Day-1 Diffusion MRI Using Multimodal Deep Embeddings for Stroke Outcome Prediction".

Jane: Here's the gist of the paper. The gist: J1 multimodal models achieved the highest predictive performance (AUC = 0.923 ± 0.085), outperforming J0-based configurations (AUC ≤ 0.86).

Tom: Study Objective This study compares baseline (J0) and 24-hour (J1) diffusion magnetic resonance imaging for predicting three-month functional outcomes after acute ischemic stroke (AIS) The goal was to determine if early posttreatment diffusion MRI provides superior prognostic value to pre-treatment imaging.

Jane: The researchers hypothesized that post-treatment (J1) imaging, combined with clinical and lesion-volume features, can improve predictive performance and interpretability.

Tom: The study analyzed seventy-four AIS patients with paired apparent diffusion coefficient (ADC) scans and clinical data.

Paper summary: Tom: What this means for us on the air is that we're seeing a clear signal that the change in diffusion within twenty-four hours after stroke is a much stronger predictor of recovery than just looking at the initial scan.

Jane: It’s interesting because they explicitly compared J0 and J1 MRI models, showing that J1 outperformed J0 across all metrics, including AUC zero point seven one four plus or minus zero point one zero five versus zero point five four zero plus or minus zero point two six three > <ref:2512.02088#pg3,J0 and J1 MRI models>

Lu: They are using the comparison of baseline versus Day-one diffusion MRI to show that the changes captured at J1 provide a stronger prognostic signal than the initial baseline scan, which is what they found in Table one > <ref:2512.02088#pg1>

Meng: It’s good to see how they built this up, integrating deep volumetric and clinical features first, and then seeing how adding those lesion-volume metrics further improves things >

Lalam: Integrating imaging features with clinical data brought the multimodal model up to an AUC of zero point eight nine four plus or minus zero point one zero six, which shows that combining the modalities is pretty valuable > <ref:2512.02088#pg3>

Conclusion: Tom: So looking at the title, Comparing Baseline and Day-one Diffusion MRI Using Multimodal Deep Embeddings for Stroke Outcome Prediction, it really highlights that the method isn't just about comparing two scans; it’s about using deep learning embeddings to build a robust prediction framework <ref:2512.02088#pg1,Comparing Baseline and Day-1 Diffusion MRI Using Multimodal Deep Embeddings for>.

Jane: And what this implies is that for stroke prognosis, we should be looking at early post-treatment diffusion MRI as a key piece of information rather than just the baseline scan >

Lu: The framework they proposed uses PCA to reduce the three dimensional ResNet-fifty embeddings down to twelve components, which makes it more stable and interpretable by ensuring >ninety-five percent variance is retained in those components >

Meng: From a deployment standpoint, having a model that can be interpreted through those principal components gives us some transparency on why the AI made a certain prediction >

Lalam: The overall conclusion is that this PCA support vector machine framework offers a clinically translatable foundation for personalized stroke prognosis by integrating quantitative diffusion MRI with structured clinical data >

Sina Raeisadigh, Myles Joshua Toledo Tan, Henning Muller, Abderrahmane Hedjoudje

Department of Computer Science, University of Geneva, Switzerland · Department of Electrical & Computer Engineering, University of Florida, FL, USA · Service of Medical Informatics, University Hospital of Geneva, Switzerland · Department of Imaging and Medical Informatics, University of Geneva

eess.IV, cs.AI, cs.CV, cs.LG

Submitted: 2025-12-01

Updated: 2025-12-01

Comments: 5 pages, 5 figures, 2 tables

Journal ref: 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), 2026, pp. 1-5

DOI: 10.1109/ISBI61618.2026.11515670

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 83/100

The gist: The gist: J1 multimodal models achieved the highest predictive performance (AUC = 0.923 ± 0.085), outperforming J0-based configurations (AUC ≤ 0.86).

Key concepts

Baseline vs. Day-1 Diffusion MRI (J0 vs J1): This compares diffusion MRI scans taken at the time of stroke (baseline, J0) with those taken 24 hours later (J1). The research tested whether the change in diffusion patterns within this first day provides a better signal for predicting long-term patient recovery than just looking at the initial scan.
3D ResNet-50 Embeddings: ResNet-50 is a deep neural network used to analyze complex 3D MRI data. It converts the detailed diffusion MRI scans into compact, meaningful numerical representations called embeddings. These embeddings capture the essential structural and tissue characteristics of the brain in a way that machine learning models can easily use.
Multimodal Fusion and PCA: This process combines different types of data—MRI features, clinical variables, and lesion volumes—into one comprehensive set. Principal Component Analysis (PCA) is then used to reduce the complexity of this combined data down to a manageable number of key components while retaining most of the important predictive information.

Terminology

AUC — 2× in this paper · explained in 2 episodes
Embeddings — 2× in this paper · explained in 2 episodes
Interpretability — 1× in this paper · explained in 15 episodes

Summary

The gist: J1 multimodal models achieved the highest predictive performance (AUC = 0.923 ± 0.085), outperforming J0-based configurations (AUC ≤ 0.86).

Study Objective

This study compares baseline (J0) and 24-hour (J1) diffusion magnetic resonance imaging for predicting three-month functional outcomes after acute ischemic stroke (AIS) The goal was to determine if early posttreatment diffusion MRI provides superior prognostic value to pre-treatment imaging. The researchers hypothesized that post-treatment (J1) imaging, combined with clinical and lesion-volume features, can improve predictive performance and interpretability.

Methodology

The study analyzed seventy-four AIS patients with paired apparent diffusion coefficient (ADC) scans and clinical data. The framework integrated 3D ResNet-derived ADC embeddings with structured clinical variables.

The key steps in the workflow included:

ADC maps were reconstructed from 1.5 T and 3 T DWI and resampled to a unified 3D resolution of 24 × 256 × 256 voxels.
Lesion-like regions were segmented using percentilebased intensity thresholds (480 and 620 ×10−6 mm2/s), followed by morphological filtering and removal of small components.
Volume features were log-transformed and concatenated with clinical data.
3D ResNet50 was used to extract embeddings from J0 and J1 ADC inputs, which were then projected to 32–256 units.
Fused multimodal embeddings were reduced using Principal Component Analysis (PCA), retaining up to 12 components >95% variance.
The final features were classified using linear support vector machines (SVM) with eight-fold stratified group cross-validation.

Key Findings on J0 vs. J1 Imaging

MRI-only models demonstrated that J1 outperformed J0 across metrics (AUC 0.714 ± 0.105 vs. 0.540 ± 0.263; accuracy 0.633 vs. 0.433; F1 0.560 vs. 0.396), indicating that J1 diffusion changes provide stronger prognostic signal than baseline. This finding confirms that early posttreatment diffusion MRI provides superior prognostic value to pre-treatment imaging.

Multimodal Performance and Lesion Volume Impact

When combining J1 MRI embeddings with full clinical vectors, the multimodal model reached AUC = 0.894 ± 0.106. Adding lesion-volume metrics derived from ADC thresholding provided additional gains, achieving AUC = 0.923 ± 0.085. The final J1 multimodal configuration (J1 + clinical + J1 lesion volume) achieved higher validation AUCs compared to the corresponding J0 configuration (J0 + clinical + J0 lesion volume; 0.923 ± 0.079 vs. 0.811 ± 0.115; Wilcoxon signed-rank test, p = 0.0078), confirming the superior prognostic value of early post-treatment diffusion MRI.

Feature Importance and Interpretability

Feature importance derived from linear SVM coefficients revealed that both clinical and MRIderived aspects contribute to outcome prediction. The highest positive weights corresponded to motor NIHSS subscores (Day 0 Right arm motor drift, Day 0 Left arm motor drift) and age. Negative weights were dominated by MRI features (e.g., MRI feat 62, 7, 77) and the NIHSS9 (Language/aphasia) subscore. This PCA–SVM pipeline offers more transparency because each principal component and SVM coefficient can be inspected, enabling clinicians to trace predictions to specific clinical or image contributors.

Conclusion

The proposed PCA–SVM framework offers a clinically translatable foundation for personalized stroke prognosis by integrating quantitative diffusion MRI with structured clinical data. The study concludes that J1 diffusion MRI provides better prognostic value compared with baseline imaging. Integrating deep MRI embeddings with clinical and lesion-volume features (AUC > 0.92) yields a robust, interpretable, and data-efficient model. PCA stabilization enhances reliability and transparency by reducing 3D ResNet embeddings to 8–12 orthogonal components >95% variance. Overall, the proposed PCA–SVM framework offers a clinically translatable foundation for personalized stroke prognosis and will be validated on multicenter longitudinal data.

Study Limitations

Study limitations include the small sample, automated lesion segmentation and scanner variability. The authors plan to test the framework in larger cohorts and explore better segmentation, radiomics and multimodality (e.g., perfusion MRI, CT angiography). The study was conducted in accordance with the principles of the Declaration of Helsinki.

References

[1] V. L. Feigin et al., “Global, regional, and national burden of stroke and its risk factors, 1990–2019: a systematic analysis for the global burden of disease study 2019,” Lancet Neurol., vol. 20, no. 10, pp. 795–820, 2021

[3] J. L. Banks and C. A. Marotta, “Outcomes validity and reliability of the modified rankin scale: implications for stroke clinical trials:

Improvements for AI systems

Improve stroke outcome prediction by incorporating J1 diffusion MRI data instead of solely relying on baseline imaging, as J1 multimodal models achieved the highest predictive performance (AUC = 0.923 ± 0.085), outperforming J0-based configurations (AUC ≤ 0.86).
Develop a robust and interpretable framework for predicting three-month functional outcomes by integrating deep MRI embeddings with clinical and lesion volume features, as combining MRI, clinical, and lesion-volume features produces a robust and interpretable framework for predicting three-month functional outcomes in AIS patients.
Enhance model stability and generalization by applying Principal Component Analysis (PCA) to reduce the dimensionality of deep embeddings to 12 components, which produced the most consistent validation AUCs while preserving interpretability for SVM classification.
Create a clinically translatable decision support system where clinicians can trace predictions to specific contributors, as Each principal component and SVM coefficient can be inspected, enabling clinicians to trace predictions to specific clinical or image contributors.
Improve feature selection by using lesion-volume metrics derived from ADC thresholding as an independent predictor, as Adding lesion-volume improves interpretability and raises AUC by 0.03, highlighting infarct size as an independent predictor.
Refine motor impairment prediction by prioritizing admission NIHSS motor subscores (Day 0 left arm motor drift and right arm motor drift) alongside age, as these had the highest positive weights in feature importance analysis.

Abstract

This study compares baseline (J0) and 24-hour (J1) diffusion magnetic resonance imaging (MRI) for predicting three-month functional outcomes after acute ischemic stroke (AIS). Seventy-four AIS patients with paired apparent diffusion coefficient (ADC) scans and clinical data were analyzed. Three-dimensional ResNet-50 embeddings were fused with structured clinical variables, reduced via principal component analysis (<=12 components), and classified using linear support vector machines with eight-fold stratified group cross-validation. J1 multimodal models achieved the highest predictive performance (AUC = 0.923 +/- 0.085), outperforming J0-based configurations (AUC <= 0.86). Incorporating lesion-volume features further improved model stability and interpretability. These findings demonstrate that early post-treatment diffusion MRI provides superior prognostic value to pre-treatment imaging and that combining MRI, clinical, and lesion-volume features produces a robust and interpretable framework for predicting three-month functional outcomes in AIS patients.