Refined AI-ASPECTS with modified atlas and lesion-load thresholds: advancing acute ischemic stroke imaging and prognostic prediction.

BACKGROUND

The artificial intelligence-assisted ASPECTS (AI-ASPECTS) system has become an increasingly common tool in clinical practice for assessing acute ischemic stroke (AIS). However, current AI-ASPECTS implementations still rely on the conventional expert-evaluation framework, which uses a simplified two-slice atlas and arbitrarily selected lesion-load thresholds. Our study aimed to develop a refined AI-assisted ASPECTS (Ref-AI-ASPECTS) framework featuring a seamless & whole middle cerebral artery (MCA) territory atlas and region-specific, optimally determined lesion-load thresholds, and comprehensively evaluate the performance of this framework across various clinical scenarios for AIS.

METHODS

We enrolled a cohort of 7,655 AIS patients from eleven centers. Modified atlas was created by expanding conventional atlas based on full MCA territory. Ref-AI-ASPECTS with modified atlas and specific lesion-load thresholds was established using a genetic algorithm. The clinical utility of Ref-AI-ASPECTS was assessed by comparing it to the conventional framework (Con-AI-ASPECTS) in terms of correlation with NIHSS scores on admission, dichotomized prediction of mRS at 3 months, and consistency with expert scoring across the training DWI data, external DWI data, expanded CT data, and real-world prospective DWI data.

RESULTS

The Ref-AI-ASPECTS frameworks with modified atlas and specific lesion-load thresholds (2% to 29%) achieved correlation coefficients (r) of -0.414/-0.438/-0.375 and AUC values of 0.665/0.723/0.707 in the training/internal validation/external validation sets, surpassing both Con-AI- (r: -0.336/-0.402/-0.331; AUC: 0.615/0.654/0.654) and expert-ASPECTS (r: -0.196/-0.206/-0.173; AUC: 0.600/0.641/0.644) (all P < 0.01). The intraclass correlation coefficients for expert- and Ref-AI-ASPECTS were 0.82 and 0.81 in the training and external validation DWI sets, respectively, exceeding those of expert- and Con-AI-ASPECTS (0.69/0.67; both P < 0.01). These improvements were consistently validated across expanded CT datasets (AUC: 0.696 and 0.679) and in a real-world prospective cohort (AUC: 0.710).

CONCLUSIONS

The Ref-AI-ASPECTS framework outperformed conventional approaches in evaluating baseline neurological deficits and predicting functional outcomes in AIS. Our findings support the potential for its wider implementation in AI-ASPECTS systems. Prospective external real‑world validation remains necessary.

TRIAL REGISTRATION: ClinicalTrials.gov Identifier: NCT04775147; chictr.org.cn Identifier: ChiCTR2400092230.