Artificial Intelligence (AI) is a trend topic in several industries, healthcare being one of them. With the raising of deep learning techniques during the last years, the capabilities of AI systems dramatically increased and several tasks considered unmanageable have been solved.
Digital Imaging and Radiology are not exempt from this transformation, and huge hype around AI has been registered in several International and Italian conferences (eg RSNA 2018, SIRM 2018). Moreover, a lot of research paper coming from Computer Science has been published about AI applications in Radiology.
A lot of attention in AI for Radiology is coming also from industry: huge companies such as Google, Microsoft, IBM are building partnership with Hospitals and healthcare research institutions to build AI based products. Reactions from radiologists are different, spanning from curiosity and interest to apprehension.
In this article we will briefly review a few limitations of deep learning, to show that there is no chance that the current technology will replace radiologists. On the other side, through a presentation of several use cases, we will show how these technology can assist radiologists, empowering workflows, reducing the risk of errors, optimizing the processes and improving the communication with the patients.
From a technical point of view deep learning brings three main advantages:
- It has best-in-class performance in a lot of applications like, for example, speech recognition, natural language processing, and image recognition. The performance improvements over competitive technologies have been so far very relevant.
- It reduces the need for feature engineering, one of the most time-consuming parts of machine learning practice.
- It provides several architectures that can be adapted to new problems relatively easily.
These advantages led to surprising results and the adoption of these technique in several commercial applications. Despite this hype around deep learning, there are still some limitations with serious impact on healthcare applications.
Firstly, deep learning is not sufficiently transparent: deep neural networks are very complex and is not possible to trace back decisions. The transparency issue, as yet unsolved, is a potential liability when using deep learning for medical diagnosis, since users might want to understand how a system made a given decision.
Moreover, deep learning cannot inherently distinguish causation from correlation. Roughly speaking, deep learning learns complex correlations between input and output features, but with no inherent representation of causality.
Finally, deep learning works well as an approximation, but its answers often cannot be fully trusted. Deep learning systems indeed are very good at some large fraction of a given domain, but they can get easily fooled.
So, for these limitations, even if several works have proven AI systems in some tasks with performances comparable to expert radiologists, we believe that clinical decisions have still to be taken by experts. Replacing radiologists with machines would presents moreover serious ethical and legal issues.
Nevertheless, we strongly believe that AI could bring huge benefit to radiology workflows. Particularly we identified 3 main areas in the radiology workflow that might be greatly improved by the adoption of AI:
- Patient Record Analysis: Natural Language Processing (NLP) tools based on deep learning can process medical reports and EPRs (Electronic Patient Records) in order to:
- provide easy accessible summary of the patient history,
- retrieve similar cases to provide additional information to radiologists.
- Radiological Image Analysis: Image Analysis tools can provide assistance to radiologists in several ways:
- computing biomarkers such as fat percentage, BMI, sarcopenia indexes, mammary density,
- counting anatomical landmarks (eg vertebrae counting).
- Reporting: again, NLP tools can help radiologists in several way:
- checking compliance to avoid errors (eg. laterality check, missing information spotting),
- assisting radiologists in structuring reporting, codification and standardization of the reports style,
- checking automatically the readability and the complexity of reports to promote a fruitful communication to patients.
The usage of these and others AI tools have the potential to:
- reduce the workload of radiologist (especially for what concerns repetitive operations),
- reduce the risks of errors,
- improve the communication with patients.
All these benefits are crucial to provide better healthcare services, especially in scenarios where hospitals need to focus on delivering high quality care, increasing services and volumes, ensuring patient satisfaction, all with financial challenges and stretched resources.