Presenter of 4 Presentations
REPRESENTATIONAL SIMILARITIES LIMIT MULTI-MODAL DEEP LEARNING IN ACUTE ISCHEMIC STROKE DECISION SUPPORT
A DEEP LEARNING ANALYSIS OF STROKE ONSET TIME PREDICTION AND COMPARISON TO DWI-FLAIR MISMATCH
Abstract
Background and Aims
DWI-FLAIR mismatch rating is an established technique to estimate onset time of ischemic stroke in cases where it’s unknown. However, identifying subtleties in imaging is challenging even for experienced raters, especially in cases close to the 4,5 hrs threshold. Based on recent successes in medical imaging, Deep Learning (DL) might realize a promising way to augment human rating and improve accuracy.
Methods
We analyzed the performance of data-efficient Convolutional Neural Networks in estimating the onset time of ischemic stroke. We developed models on DWI and FLAIR imaging (N=489) and utilized unlabeled image data for pre-training (N=609). We tested a potential decision support scenario by augmenting human ratings with DL predictions in undeterminable cases. Additionally, we conducted an analysis to gain insights about DL predictions by post-hoc gradient-based explanations.
Results
Our DL approach improved the sensitivity of junior rating from 0.33 to 0.50, and senior rating from 0.43 to 0.48 while keeping specificity on a high level of 0.83. Junior rating gained more from augmentation by DL predictions than by senior rating due to notably high sensitivity of the model in undeterminable cases. Post-hoc explainability showed that our DL model successfully derived an association between stroke lesion and onset time without the need of explicit delineation.
Conclusions
We showed a potential use case where DL and post-hoc explanations could increase efficiency of stroke treatment selection in cases of patients with unknown onset times. Our approach does not depend on advance imaging or manual delineation, facilitating acceleration of acute decision making and aiding inexperienced raters.
ARTIFICIAL INTELLIGENCE FOR DECISION SUPPORT IN ACUTE ISCHEMIC STROKE CARE: A SYSTEMATIC REVIEW
Abstract
Background and Aims
In recent decades, several biomarkers have emerged from prospective trials facilitating the time-sensitive decision making in ischemic stroke. Despite general successes, they lack personalization, which has prompted attempts to revolutionize stroke diagnosis and treatment decision support through artificial intelligence (AI) methods. We created a systematic review of decision support systems using AI for acute ischemic stroke care that aims to provide an overview of the level of methodological robustness and standardization within the field.
Methods
We included full-text publications that propose an AI method for decision support in adult patients with acute ischemic stroke in the acute setting. Our primary objective was to describe the data, methodology, performance and outcomes used in those systems. A total of 139 studies met our inclusion criteria. Among these, 65 were included for full extraction while 74 studies were identified as proposing a method for automated stroke scoring only.
Results
There was a high degree of heterogeneity in the data sources used, methods applied, and reporting practices that were followed in our sample. We found a low level of adherence to the MINIMAR checklist.
Conclusions
There has been undeniable progress in the performance of AI models, however our results suggest significant potential validity threats, dissonance in reporting practices and challenges to clinical translation across the studies we reviewed. We outline practical recommendations for clinicians, researchers, journals and funding agencies, who all have a role to play in safeguarding the quality of research and in enabling the exploitation of the promise of AI in acute ischemic stroke care.
REPRESENTATIONAL SIMILARITIES LIMIT MULTI-MODAL DEEP LEARNING IN ACUTE ISCHEMIC STROKE DECISION SUPPORT
Abstract
Background and Aims
Treatment decisions in stroke are based on information from multiple sources. With the rise of Deep Learning (DL) based decision support, it is desired that systems emulate this ability of medical experts and aggregate multi-modal information to reach clinically feasible decisions. However, there is limited published research in this direction and proposed solutions struggle to outperform unimodal variants. This suggests that despite the practical advantage of DL with respect to seamless integration of imaging and clinical data, networks might learn redundant rather than complementary representations from different modalities.
Methods
We constructed a potential use case by using DWI and FLAIR imaging combined with clinical variables and trained multi-modal DL models for predicting outcome. Measure of similarity was computed between learnt representation of the different inputs by using Orthogonal Procrustes, a state-of-the-art technique to compare Neural Network representations. The contribution of modalities in predictions were tested by adapted feature importance rating.
Results
Multi-modal prediction performance was 0.854 AUC. We observed significantly high similarity between representations learnt from DWI vs. FLAIR imaging (similarity coefficient of 0.963) and image modalities vs. clinical data (DWI=0.615, FLAIR=0.610). Feature importance rating revealed that predictions were mainly based on clinical data, contribution of imaging was marginal.
Conclusions
Our results suggest that DL models do not necessarily learn complementary information from multiple data sources and might not exploit the different imaging traits in different sequences. Our work provides a methodological framework and an important first step towards understanding limitations of multi-modal DL and advancing solutions for better decision support in stroke.