How will AI change the practice of IR?

A longer, more scientific post after the break last week! It is a popular notion that diagnostic radiology will be greatly changed with the advent of artificial intelligence (AI), but here I discuss some interesting ways interventional radiology can benefit from AI as well.

 

AI is an exciting prospect for the entire medical field, and with radiology being at the forefront of new innovations, radiologists hold high expectations of how it will revolutionise daily practice. For interventional radiology (IR) specifically, there are two main areas where the use of AI technology can change interventional practice.




First is the pre-procedural identification of appropriate patient groups to undergo interventions. This can be achieved by harnessing the analytical prowess of AI. It has been shown that a deep learning neural network is able to achieve comparable diagnostic accuracy of intracranial aneurysms when interpreting 3D time-of-flight magnetic resonance angiography (TOF-MRA) images compared to expert radiologists [1]. The rapid diagnostic turnover will be particularly helpful for time-critical procedures such as mechanical thrombectomy in stroke patients. Machine learning algorithms can be trained to consider clinical information and imaging findings to synthesise not only a diagnosis, but to propose suitable interventions taking into account projected short/long term complication rates and estimated overall prognosis. This will enhance the decision-making process regarding whether procedural or conservative management would be more appropriate, and will provide objective data for patients to make an informed choice regarding their preferred option. On a more simplistic level, it can serve as a risk stratification tool for triaging patients to undergo interventions. For example, by determining whether breast tissue biopsy is indicated based on mammographic findings alongside patient demographics, family history and other relevant risk factors.


The other potential utility of AI in IR practice is by enhancing intra-procedural accuracy. AI can aid anatomical recognition of the relevant structures. This will be complemented by the improvement in existing diagnostic modalities and the development of new technology. Looking towards the future, AI may be able to analyse the anatomy of a tortuous vessel that is difficult to navigate with a catheter, and overlay a highlight over the key landmarks using augmented reality [2]. This will help cut down on procedure duration along with complication rates. As an extended benefit, this could reduce radiation dose to the patient and the radiologist by minimising back-and-forth manoeuvring and re-imaging. Utilisation of robotics in current procedures is a way to increase reproducibility and reduce operator dependency. The ethico-legal implications aside, it is within the realms of possibility for AI to eventually assist or autonomate robotics with a high degree of precision.



Furthermore, AI will also improve IR through non-clinical means. It may be used by IR trainees to learn to cope under pressure by running realistic simulations of procedures based on data of actual cases fed into the algorithm. It may be used to streamline quality improvement processes by automatically running analysis of various outcome measures and comparing them to the standard. The possibilities are endless.


In conclusion, I believe that AI will positively impact IR practice and it will be essential for all clinicians to adapt and evolve alongside AI.



References


1. Faron A, Sichtermann T, Teichert N, Luetkens JA, Keulers A, Nikoubashman O, Freiherr J, Mpotsaris A, Wiesmann M. Performance of a Deep-Learning Neural Network to Detect Intracranial Aneurysms from 3D TOF-MRA Compared to Human Readers. Clin Neuroradiol. 2020 Sep;30(3):591-598. doi: 10.1007/s00062-019-00809-w. Epub 2019 Jun 21. PMID: 31227844.

2. Midulla M, Pescatori L, Chevallier O, et al. Future of IR: Emerging Techniques, Looking to the Future…and Learning from the Past. J Belg Soc Radiol. 2019;103(1):12. Published 2019 Jan 28. doi:10.5334/jbsr.1727

11 views0 comments

Recent Posts

See All