Stroke outcome and evolution prediction from CT brain using a spatiotemporal diffusion autoencoder

Adam Marcus; Paul Bentley; Daniel Rueckert

cs.CVFeb 28, 2026

Stroke outcome and evolution prediction from CT brain using a spatiotemporal diffusion autoencoder

Authors: Adam Marcus, Paul Bentley, Daniel Rueckert

Organizations: Imperial College London, London, UK · Technische Universität München, München, Germany

Abstract

Stroke is a major cause of death and disability worldwide. Accurate outcome and evolution prediction has the potential to revolutionize stroke care by individualizing clinical decision-making leading to better outcomes. However, despite a plethora of attempts and the rich data provided by neuroimaging, modelling the ultimate fate of brain tissue remains a challenging task. In this work, we apply recent ideas in the field of diffusion probabilistic models to generate a self-supervised semantically meaningful stroke representation from Computed Tomography (CT) images. We then improve this representation by extending the method to accommodate longitudinal images and the time from stroke onset. The effectiveness of our approach is evaluated on a dataset consisting of 5,824 CT images from 3,573 patients across two medical centers with minimal labels. Comparative experiments show that our method achieves the best performance for predicting next-day severity and functional outcome at discharge.

Figures & tables

Figure 1: Overview of the (A) spatial and (B) spatiotemporal diffusion autoencoder approaches. In both cases, a semantic encoder takes an image containing a stroke lesion and maps it to a latent code. A Denoising Diffusion Probabilistic Models (DDPM) [ 18 ] is then conditioned on this latent code to denoise a different image of the same lesion taken either at (A) the same time or (B) a future point in time. For the spatiotemporal method, the latent code is also concatenated with the future time using a multilayer perceptron (MLP). After training, the semantic encoder can then be fine-tuned with minimal data and used to predict a stroke outcome.

Parameter	Value
Base channels	16
Channel multipliers	[1, 2, 4, 8, 16, 32, 64]
Attention resolution	[16]
Encoder base channels	16
Encoder attention resolution	[16]
Encoder channel multipliers	[1, 2, 4, 8, 16, 32, 64]

Table 1: Network architecture of our diffusion model based on the improved Denoising Diffusion Probabilistic Model (DDPM) of Dhariwal et al. [ 11 ] .

Characteristic	Train and validation set (n = 2858)	Test set (n = 715)
Number of CT, mean (range)	1.63 (1–5)	1.63 (1–5)
Missing outcome, n (%)	1178 (41.2%)	276 (38.6%)
Age (years), median (IQR)	75.0 (62.4–83.7)	75.0 (65.0–83.0)
Female sex, n (%)	914 (54.4%)	242 (55.1%)
ASPECTS, median (IQR)	10 (9–10)	10 (9–10)
NIHSS on admission, median (IQR)	7 (4–13)	7 (4–13)

Table 2: Population characteristics of the clinical dataset.

Method		AUC	ACC	F1	AUC	ACC	F1	FID	MSE
		24-hour NIHSS			Discharge mRS
CNN	Direct training	0.584	59.9	52.0	0.702	60.4	57.4	—	—
	VICReg [ 4 ]	0.582	59.9	52.0	0.711	63.1	58.6	—	—
Autoencoder	Variational [ 20 ]	0.628	62.9	63.6	0.726	64.4	63.5	8.9	69.9
	Diffusion [ 26 ]	0.623	62.6	64.2	0.735	65.8	64.2	7.1	47.7
Ours	Spatial	0.648	63.6	65.4	0.757	67.7	69.2	7.3	47.9

Table 3: Stroke outcome prediction results obtained by our method and spatiotemporal ablation variants compared to baseline approaches. In our method, a pair of 2D image slices,

\mathbf{x}_{a}

and

\mathbf{x}_{b}

, showing the same stroke lesion are sampled, where

\mathbf{x}_{a}

is always from the earliest scan and the DDPM receives a noisy variant of

\mathbf{x}_{b}

taken at a later time. In the Any forward pair variant,

\mathbf{x}_{a}

is from any time prior to

\mathbf{x}_{b}

, and in the Any pair variant,

\mathbf{x}_{a}

and

\mathbf{x}_{b}

can be views from any time. For fair comparison all methods utilized a ResNet-50 [ 14 ] as the semantic encoder and followed an identical fine-tuning procedure, unless trained directly.

Figure 2: Example reconstructed image of a right middle cerebral artery (MCA) stroke from our test set for different methods. The performance of our spatiotemporal approach and the diffusion autoencoder (AE) are similar and superior to a variational AE.

Explore similar work

StrokeNeXt: A Siamese-encoder Approach for Brain Stroke Classification in Computed Tomography Imagery
Feb 16, 2026Leo Thomas Ramos, Angel D. SappaMedical Image Classification Classification
Computer Vision Center · Universitat Aut`onoma de Barcelona · ESPOL Polytechnic University
Bi-temporal Image-driven Acute Stroke Evolution Analysis
Feb 7, 2026Md Sazidur Rahman, Kjersti Engan, Kathinka Dæhli Kurz +1Longitudinal Medical Image Analysis Computed Tomography
Department of Electrical Engineering and Computer Science, University of Stavanger, Norway · Stavanger Medical Imaging Laboratory, Stavanger University, Hospital, Norway
StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-contrast CT
Jun 3, 2026Weiru Wang, Susanne G. H. Olthuis, Elizaveta Lavrova +4Medical Image Analysis Medical Image Representation Learning
Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands · Graduate School of Life Sciences, Utrecht University, Utrecht, The Netherlands · Department of Neurology, Maastricht University Medical Centre+, Maastricht, The Netherlands +3

Feb 16, 2026J/K move · Enter open · S save

StrokeNeXt: A Siamese-encoder Approach for Brain Stroke Classification in Computed Tomography Imagery

Leo Thomas Ramos, Angel D. Sappa

Computer Vision Center · Universitat Aut`onoma de Barcelona · ESPOL Polytechnic University

Medical Image Classification Classification Convolutional Neural Networks Computed Tomography Siamese Neural Networks

Open paper AskPDF arXiv

We present StrokeNeXt, a model for stroke classification in 2D Computed Tomography (CT) images. StrokeNeXt employs a dual-branch design with two ConvNeXt encoders, whose features are fused through a lightweight convolutional decoder based on stacked 1D operations, including a bottleneck projection and transformation layers, and a compact classification head. The model is evaluated on a curated dataset of 6,774 CT images, addressing both stroke detection and subtype classification between ischemic and hemorrhage cases. StrokeNeXt consistently outperforms convolutional and Transformer-based baselines, reaching accuracies and F1-scores of up to 0.988. Paired statistical tests confirm that the performance gains are statistically significant, while class-wise sensitivity and specificity demonstrate robust behavior across diagnostic categories. Calibration analysis shows reduced prediction error compared to competing methods, and confusion matrix results indicate low misclassification rates. In addition, the model exhibits low inference time and fast convergence.

Abstract

Explore similar work

StrokeNeXt: A Siamese-encoder Approach for Brain Stroke Classification in Computed Tomography Imagery

Bi-temporal Image-driven Acute Stroke Evolution Analysis

StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-contrast CT