Page 13 - AIH-2-4

Artificial Intelligence in Health AI in acute stroke imaging

3.4. Automation: Workflow and triage 3.6. Ethical, legal, and social implications of AI in

Commercially available algorithms are being integrated stroke imaging
into the clinical workflows of numerous large institutions, The application of AI has revolutionized stroke imaging;
both in practice and trials, to provide automated triage however, ethical, legal, and societal implications present
and segmentation of acute stroke cases. These tools barriers that need to be addressed. Potential biases in
help decrease the workload on radiologists and enhance training data and the decision-making process of AI (often
diagnostic accuracy. Automated ASPECT scoring supports referred to as the “black box” nature) raise ethical and
treatment teams in selecting patients for endovascular societal concerns. These can be mitigated by implementing
therapy. Overall, such tools offer rapid and efficient a robust framework that emphasizes data security, patient
analyses to improve stroke care at both spoke and hub privacy, and fair and equitable access to AI applications in
hospitals and reduce the turnaround times in medical healthcare. 73
workflows. In a retrospective study by Colasurdo et al., Multidisciplinary discussions on the advantages and
59
a convolutional neural network was incorporated into the limitations of using AI in healthcare, among all stakeholders,
institutional workflow to detect SDH from NCCT head
scans. The subdural convolutional neural network showed including clinicians, AI developers, administrative
91.4% sensitivity, 96.4% specificity, and 95.1% accuracy; personnel, and policymakers, are essential. Standardized
for the subgroup with SDH thickness greater than 10 mm, protocols and regulations should be established to promote
the sensitivity reached 100%. impartiality, clarity, trustworthiness, accountability,
confidentiality, and compassion in the development of AI
Retrospective research by Soun et al. also within an ethical framework. 74
1
demonstrated that the integration of an AI algorithm into
the hospital system supported triage of CTA in ischemic 4. Challenges and future directions
stroke cases, enabling automatic identification of LVO The development and deployment of AI platforms in
cases with 96% sensitivity and 85% specificity, and a clinical settings have been instrumental in transforming
turnaround time of 22 min. The AI was accessible in the stroke care by lowering mortality and improving quality
Picture Archival and Communication System via a web of life. However, several challenges constrain their
or mobile application.
widespread adoption. One major challenge is the limited
3.5. Integration with teleradiology workflow generalizability of datasets, i.e., AI models are often
trained on single-center or homogeneous datasets, which
Teleradiology has been addressing the global challenge may result in underperformance when applied to external
of radiologist shortages. 60-64 The seamless integration of populations or systems with different scanners, imaging
cloud-based AI solutions with telereporting platforms protocols, electronic health record systems, laboratory
enhances workflow by prioritizing critical cases, sharing equipment, and varying clinical and administrative
automated alerts to stroke teams for prompt action, and procedures. To improve generalizability and performance,
extending the benefits of AI across multiple domains, continuous learning from large, diverse, multicenter, and
including remote or underserved areas. 65-71 However, data high-quality annotated datasets is essential. 29,75
security poses a significant challenge for cloud-based AI
algorithms. Implementing robust cybersecurity systems Another challenge is the “black box” nature of numerous
is pivotal to ensure secure integration of AI into the AI models, which limits interpretability, reliability,
teleradiology workflow. Another challenge is ensuring and transparency in their decision-making processes,
that AI outputs are accessible within the teleradiologist’s hindering widespread acceptance (1). The development
viewer. The use of aggregator platforms and workflows that and implementation of heat maps, prediction-based
consolidate outputs from multiple AI tools would support modules, user-friendly interfaces, interactive dashboards,
seamless telereporting. Furthermore, leveraging clinical and visualization tools can help make AI insights more
data from teleradiology, incorporating feedback from understandable, thereby addressing the “black box”
teleradiologists, and collaborating with AI developers for problem. The white paper of the Italian Society of Medical
76
training, upgrading, and validating AI systems will further and Interventional Radiology emphasizes the urgent need
streamline integration in real time. A potential obstacle for explainable AI (xAI), which can reveal the rationale
is the lack of adequate infrastructural support for AI behind AI decision-making, offering insights into its
integration. Employing Graphic Processing Units would strengths, limitations, and potential future performance.
77
allow efficient and fast processing of large volumes of data Furthermore, the ongoing training of technologists,
for AI development. 13,72 radiologists, and physicians through workshops and

Volume 2 Issue 4 (2025) 7 doi: 10.36922/AIH025140025