Abstract
Background: Although artificial intelligence systems that diagnosis among different conditions from medical images are long term aims, specific goals for automation of human-labor, time-consuming tasks are not only feasible but equally important. Acute conditions that require quantitative metrics, such as acute ischemic strokes, can greatly benefit by the consistency, objectiveness, and accessibility of automated radiological reports. Methods: We used 1,878 annotated clinical brain MRIs from patients with acute ischemic strokes to generate a fully automated artificial intelligence based system that outputs radiological reports in addition to the infarct volume, 3D digital infarct mask, and the feature vector of anatomical regions affected by the acute infarct. Findings: The performance of our system was comparable to that of an expert evaluator. The weight of the components of the feature vectors that supported the prediction of the reports, as well as the prediction probabilities are outputted, making the pre-trained models behind our system interpretable. The system is publicly available, runs in real time, in local computers, with minimal computational requirements, and it is readily useful for non-expert users. Interpretation: Our system supports large-scale processing of new and legacy data, enabling clinical and translational research. The generation of reports indicates that our fully automated system is able to extract quantitative, objective, structured, and personalized information of potential clinical relevance from stroke MRIs.
Introduction
In the high-stakes environment of acute ischemic stroke, "time is brain." Clinicians require precise quantitative metrics—such as the exact volume of infarcted tissue and the specific anatomical regions affected—to make critical treatment decisions. Currently, generating these reports is a manual, labor-intensive process prone to inter-observer variability and difficult to scale for large-scale research. While many AI tools focus on simple binary detection, there is a significant gap in tools that provide the structured, detailed anatomical descriptions required for clinical practice.
To address this, researchers developed ADSv1, a fully automated AI pipeline that transforms raw acute stroke MRIs into comprehensive, expert-level radiological reports in seconds.
The ADSv1 Pipeline
The Automated Detection and Segmentation of Acute Stroke (ADSv1) system utilizes a multi-step workflow to convert raw imaging data into actionable clinical insights.
As illustrated in the pipeline overview, the process involves:
- Segmentation: A deep-learning algorithm detects and segments the ischemic core from Diffusion-Weighted Imaging (DWI) and Apparent Diffusion Coefficient (ADC) maps.
- Standardization: The patient’s brain is mapped to a common template (MNI space) using 12-parameter linear deformation.
- Feature Extraction: The system calculates Quantitative Feature Vectors (QFVs). These represent the ratio of injury within specific Regions of Interest (ROIs) based on arterial territories (e.g., MCA, ACA) and classical anatomy (e.g., basal ganglia).
- Classification and Interpretability: Using Random Forest models and SHAP (SHapley Additive exPlanations) values, the system predicts infarct locations and provides transparent reasoning for its findings, ensuring the model is not a "black box."
Performance and Validation
The study utilized a robust dataset of 1,878 annotated clinical MRIs from the Johns Hopkins Comprehensive Stroke Center. The data represented a wide variety of technical conditions, including 11 different scanners and both 1.5T and 3T magnetic fields, ensuring the model's robustness to real-world hardware variations.
Key results include:
- Expert-Level Accuracy: The system achieved excellent agreement with expert visual analysis, with a Balanced Accuracy (BACC) and F1 score > 0.8 for the majority of brain regions.
- Efficiency: The system runs in real-time on local computers, outputting reports in seconds.
- Demographic Robustness: No significant differences in prediction accuracy were found regarding the patient’s sex, race, or the magnetic field strength of the scanner.
- Clinical Utility: The system successfully predicted secondary complications like hydrocephalus (ventricular enlargement or midline shift) with a BACC of 0.84.
Discussion and Future Directions
A central contribution of this work is the creation of "computable data objects." By outputting structured QFVs rather than just narrative text, ADSv1 allows imaging data to be immediately ingested by other AI models or used in large-scale functional studies. This bridges the gap between clinical imaging and data science.
While the system shows high performance, challenges remain in regions where infarcts are statistically rarer (such as the ACA territory) and in cases of extreme brain distortion where linear mapping may struggle. Future work will focus on increasing dataset diversity and refining automated registration for highly distorted brains. ADSv1 is now publicly available, providing a tool for both clinical decision support and large-scale translational research.
Summary updated