Artificial Intelligence in Radiology: A Private Practice Perspective From a Large Health System in Latin America

Section snippets

Dipping one's Toes in the Water of AI

Getting started with AI can be challenging. It is important to become familiar with the basic terms in machine learning to understand the purpose of each type of algorithm and the intended use of specific algorithms.

In a nutshell, an algorithm's primary goal can be to detect, classify, segment, and/or quantify something. Depending on the objective, a simple detection algorithm can identify the most critical cases, reprioritize the worklist and reduce the turnaround time for reporting.

In

The Data

Whether validating third-party software or creating one's own, data is an essential part of any artificial intelligence project.

When evaluating a vendor's solution, there are some key considerations to help determine whether the algorithm presented is reliable:

• 1Exploring external validation articles published in peer-reviewed journals.
• 2Talking to other institutions that used the model and ask for their feedback.
• 3Performing an external validation with one's own data.

We strongly recommend the third

Data Security

When working with patient-protected health information, an important step is related to sensitive data protection. During the anonymization process, personal/protected health information (PHI) contained in the Digital Imaging and Communications in Medicine (DICOM) metadata must be replaced by random, hashed, or empty string, making it difficult to re-identify the patient. Special care must also be given to the image itself. Especially in X-Ray, mammograms, and ultrasound images, it is common to

Develop or Buy

Most AI providers work on a pay-per-use model, in which case payment is done according to the number of inferences. Others propose fixed models according to the number of equipment, for example. It is important to analyze which strategy works for each institute or health system. Alternatively, it is worth trying to develop the algorithm internally for institutions where such capability exists. It is useful to calculate the net present value, the internal rate of return, or other metrics related

Project Selection

According to the IBM Global AI Adoption Study, 35% of companies reported using AI algorithms and, an additional 42% reported exploring AI.⁸ Along with this high interest in AI, comes high expectations of returns.

Therefore, the success of an AI project starts with a fundamental question. Which project to choose? It is up to us, experts in the field, to know the AI market and identify all solutions that exist. In addition, we need to analyze our company's pains and identify which of them can be

Computer Vision Projects

While disease detection projects are a no-brainer first project option, it is not always easy to prove return on investment. For these projects, it is important to constantly monitor results, computing performance metrics between the algorithm output and the radiologist's opinion. An option is to measure how the algorithm helped reduce errors. It is also helpful to analyze how the reduced turnaround time improved the outcome or made it possible to quickly discharge patients with normal exams,

Natural Language Processing Projects

Although these projects do not attract as much attention as computer vision algorithms, we have seen great success with them. The evolution of deep learning techniques, and especially the transformer architecture, has increased the performance of these algorithms. Considering that all imaging studies, as well as all medical care generate texts, it can be valuable for companies to analyze this rich, but often unstructured, data.

It is highly recommended that radiologists notify critical findings

Monitoring and Feedback

Machine learning algorithms are trained using finite data, and various techniques are used to increase the model's ability to perform well even on never-seen-before data, which is referred to as generalization. However, it is common to observe the performance worsening over time. These drifts can occur for several reasons, such as a change in the population, in the way data is generated, or even in disease distribution, as we saw during the COVID-19 pandemic.

That is why it is essential that we

Bias and Fairness

Within the context of real-world applications, great attention must be paid to ensure fairness in the use of AI solutions. As seen earlier, failures in the project design, data selection, or training can generate algorithms that are not representative of a given population.

Gichoya et al.¹⁰ have shown that models are able to identify race in a simple chest X-ray. This exposes how some factors invisible to the naked eye can impact a deep learning model. To reduce flaws in deep learning models,

Final Thoughts

Artificial intelligence in radiology is already a reality, and its use is becoming more widespread. Although several algorithms have demonstrated improved quality, their acquisition, implementation, and monitoring can be costly. Within a private service, in addition to improving quality and safety, the choice must also consider the return on investment.

It is important to know the solutions in the market, identify the institution's practical challenges and assess which ones are a priority and

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

We would like to thank Dasa for its support investing in the development of AI in the healthcare. We thank Romeu Domingues, Emerson Gasparetto, Leonardo Vedolin, Bernardo Bizzo, Marcio Garcia, Gustavo Pinto, Alexandre Valim, Roberto Cury, Renata Brunelli, Hommenig Scrivani, and Pedro Bueno for the vision, the creation, the execution, and the partnership with the AI Lab. We thank all the people who work at the AI Lab and built this journey. And our sincere gratitude to Dr. Reza Forghani of the

‍