The Future of Artificial Intelligence in Diagnostic Imaging


New AI models open new frontiers and promise to revolutionize Diagnostic Imaging. The FDA has approved more than 200 commercial AI products for Radiology. Hospitals and healthcare facilities already use them successfully, but some obstacles must be overcome before widespread clinical use can be appreciated.

Artificial intelligence (AI) models have been successfully used in many fields of science and technology and, with the promise of greater effectiveness and efficiency, have established themselves in the vision of companies and institutions.

One of the most important research advances in the field of AI has occurred in the field of Computer Vision (CV). This area focuses on the analysis and interpretation of images and videos.

Artificial Intelligence at the Center of Diagnostic Imaging

The medical discipline of Diagnostic Imaging has found itself at the center of a scientific and productive fervor that has all the prerequisites to lead to a disruption in the interpretation of medical images.

AI in Radiology has shown great success in detecting and classifying abnormalities on radiographs, computed tomography (CT), and magnetic resonance imaging (MRI) scans, leading to more accurate diagnoses and greater efficiency in decision-making.

An article recently published in “The New England Journal of Medicine (NEJM)” aims to illustrate the current situation and future scenarios of using Artificial Intelligence techniques in Diagnostic Imaging.

AI in Radiology: PACS, RIS, and DICOM

Compared to other fields of medicine, radiology is in a favorable position to adopt AI algorithms and infrastructures.

The key factors are the presence of an established digital workflow and universal standards for image archiving. Think of PACS (Picture Archiving and Communication System), RIS (Radiology Information System), and the DICOM (Digital Imaging and Communications in Medicine).

PACS (Picture Archiving and Communication System) is a radiological image archiving and communication system. It consists of hardware and software that allows you to acquire, store, view, and share diagnostic images generated by medical imaging devices.

PACS eliminates the need for traditional X-ray films, allowing digital and centralized image management. This makes it easier for doctors to view and share images between healthcare facilities.

RIS (Radiology Information System) is a radiological information system that manages the administrative and clinical activities of the radiology department.

RIS records and manages information such as patient appointments, radiology exam scheduling, report archiving, and integration with PACS.

It provides:

  • A central platform for managing radiology data.
  • Facilitating exam scheduling.
  • Report generation.
  • Communication between medical team members.

DICOM (Digital Imaging and Communications in Medicine) is an international standard for acquiring, storing, transmitting, and displaying medical images.

This standard defines a common format and data structure for digital radiology images, enabling interoperability between different imaging systems and devices.

DICOM allows you to associate clinical information with image data, such as medical reports, acquisition parameters, and other relevant information. It is essential to ensure that radiological images can be easily shared and interpreted correctly by different healthcare professionals and IT systems in the medical context.

The aforementioned standardized systems offer a great opportunity, providing a solid structural basis for AI integration.

Artificial Intelligence and Diagnostic Imaging in Early Diagnosis

Radiological algorithms can improve preliminary image processing processes, including image acquisition, reconstruction, and noise mitigation; they can also perform various functions to assist radiologists.

Artificial Intelligence, for example, is used for the detection, localization, and classification of conditions such as lung nodules and breast anomalies (this is the case, for example, in our country, of the AI-based device used at the IRCCS Polyclinic of Sant’Orsola to support the early diagnosis of lung nodules).

Tools of this type have interpretative capabilities that can sometimes exceed those of expert operators, particularly in predicting clinical outcomes using digital biomarkers.

AI in Radiology: satisfied but also worried radiologists

AI in radiology has attracted global interest, and commercial algorithms are now developed by companies with operations in more than 20 countries.

The Food and Drug Administration (FDA) has approved more than 200 commercial artificial intelligence products for radiology.

While some hospitals and treatment centers are already successfully using these products, substantial hurdles must be overcome before widespread and successful clinical use can be seen.

Radiologists who use AI in their clinical practice are generally satisfied with their experience and report how it provides value to them and their patients. However, concerns

arise caused by a lack of knowledge and trust and possible changes in professional identity and autonomy.

Among the elements that can help overcome the concerns above:

  • Presence of AI expert colleagues
  • Education and Training
  • Support

A co-pilot role in the future of AI in Radiology

Most radiologists expect substantial changes in the profession within the next decade and believe that AI should be a “co-pilot,” acting as a second reader and improving workflow management.

Although the presence of AI in the US market is estimated at only 2%, radiologists’ attention and the technology’s potential indicate that further progress in adoption into clinical practice is likely in the near future.

Artificial Intelligence in Diagnostic Imaging also for non-radiologist doctors?

The use of AI in radiology is wider than radiologists. There is an emerging global trend toward using these tools by non-radiologist clinicians and other healthcare professionals.

This could improve access to medical imaging and reduce diagnostic errors in low-resource settings and emergency departments.

AI techniques can accelerate the acquisition of medical images outside of traditional reference workflows, for example, through portable tools such as Swoop (portable magnetic resonance imaging controllable via a tablet) or ultrasound probes connected to smartphones.

Although not well established, this type of use has been cited as a potential long-term threat to Radiology as a specialty because advanced AI models can reduce the complexity of technical interpretation so that a non-radiologist physician can use imaging without relying on a trained radiologist.

However, this emerging trend of using radiological image interpretation algorithms outside their usual reference domain may allow for greater accessibility and ultimately improve patient health.

AI in the clinical field: the obstacles to overcome

There are still some obstacles to overcome before we can see a wider adoption of AI in the clinical setting, including the challenge of generalizability: the presence of such platforms in Radiology raises the question of their effectiveness for all typologies of patients, as the underlying models are not often tested outside of the environment in which they were developed.

If, on the one hand, there is a lack of controlled clinical trials, on the other, many radiological models deteriorate in performance when applied to patients other than those used for model development, a phenomenon known as “data set shift.”

A ubiquitous presence of such tools in Radiology requires validation guarantees that include greater collaboration between clinicians and AI and promises of transparency and monitoring to improve the generalization of algorithms in new contexts.

At the basis of the diffusion and correct use of AI is the understanding and implementation of the respective strengths and weaknesses of the human operator and the technological tool: similarly to what happened in the field of chess, the best performance seems to be the one produced by an adequate collaboration between the two, even if many studies have reported uncertain results.

While collaboration between clinicians and AI can improve the diagnosis and treatment process, more transparency in AI models can help clinicians’ understanding and trust.

The main open problem is, perhaps, the models’ interpretability, which makes clinicians reluctant to use black boxes whose reasoning behind the outputs they cannot analyze.

Transparency is also a problem in evaluating the generalization of AI algorithms in the medical field since an independent evaluation of the models needs to be included. The reports of the tools approved by the regulatory authorities often need more fundamental information, such as sample size, characteristics, demographics, and device specifications.

AI models: how to make them reliable for clinical practice 

The aforementioned NEJM article cites some plausible solutions to the problems identified so far, for example:

  • The use of specific checklists
  • The curation of public datasets of medical images as benchmarks for the effectiveness of AI products (with resulting privacy issues, high costs, and overrepresentation of specific patient populations)
  • Federated learning opens up the possibility of training models in a decentralized manner.
  • The sharing of highly heterogeneous datasets to improve the generalizability of the results.

Monitoring instruments in clinical practice can improve accuracy and reliability, but regulatory restrictions may limit updating models after approval.

This implies that changes in disease prevalence and incidence, advances in medical technology, and clinical practice need to be reflected in adequate model performanceTherefore, continuous updating through clinicians’ feedback and techniques such as continuous learning (continuous Training of the model) is essential.

Artificial Intelligence and Diagnostic Imaging: Towards Generalist Models

The current generation of Radiology models mainly focuses on a limited set of interpretation tasks, such as detecting specific lesions or classifying certain pathologies.

However, this approach only partially reflects the complexity and variety of tasks that radiologists face on a daily basis. As a result, many Radiology professionals were initially skeptical about adopting Artificial Intelligence (AI) techniques into their practice.

However, a new generation of generalist medical AI models is developing and could change the radiological landscape.

These generalist models are designed to address the task of interpreting radiological images in their entirety, covering a wide range of pathologies and medical conditions. Their potential ability to assist the entire diagnostic and therapeutic process could significantly evolve radiology practice.

Generalist AI models in Cardiology and Oncology

Areas, where these generalist AI models could have a significant impact, include early stroke identification and cancer detection.

Thanks to their ability to analyze large amounts of radiological data, these models can detect early signs of stroke and provide a timely diagnosis, initiating rapid and effective therapeutic interventions.

Specifically, they can aid in detecting brain hemorrhages, blood vessel blockages, or ischemic brain tissue, making the work of radiologists easier and improving the clinical outcome for stroke patients.

AI “discovers” new pathologies without Training

Furthermore, recent developments in AI have led to new possibilities, such as “zero-shot learning.” This approach allows models to learn new pathologies or conditions without specific Training. In practice, a model could be trained on a set of radiological images relating to specific pathologies but would subsequently be able to recognize and correctly interpret other pathologies never encountered during the training phase. This flexibility and ability to adapt represents a significant advance in the implementation of AI in radiology.

New frontiers: multimodal and self-supervised models

Another promising frontier is represented by multimodal models, which integrate the analysis of radiological images with other modalities of clinical data, such as medical texts and reports. This integration allows models to understand complex clinical contexts and provide more accurate case assessments by combining information from different data sources.

Likewise, self-supervised models are emerging as a promising methodology, where models learn to interpret radiological images without the need for explicit labels or annotations, using intrinsic information

present in the data itself.

The Future of Artificial Intelligence in Diagnostic Imaging

The progress of the Radiology industry is set to be significantly accelerated by the development and implementation of Artificial Intelligence (AI) models.

The advancement towards generalist medical AI models, capable of processing imaging data, speech, and medical text and generating outputs that reflect advanced medical reasoning, opens up new perspectives for radiology practice.

Future AI models can provide detailed natural language explanations, spoken cues, and image annotations that reflect sophisticated medical reasoning. They can generate customized text output based on the specific characteristics of medical images, answering end users’ needs and questions more accurately and effectively. This level of personalization and contextual understanding will enable clinicians to obtain high-quality information and support during their radiology practice.

AI is destined to revolutionize Healthcare, but man remains at the center

Integrating generalist medical AI models into Radiology promises to transform this discipline and the entire healthcare industry.

These AI models could help radiologists improve the accuracy and efficiency of diagnoses, enabling more timely assessment and optimized disease management. Furthermore, they could provide valuable guidance for treatment planning, helping doctors make informed and personalized decisions.

Overall, the potential of generalist medical AI models to provide comprehensive solutions in the field of Radiology is set to transform this discipline and Healthcare as a whole.

Their impact could extend beyond Radiology, impacting other medical specialties and improving the overall effectiveness of patient care.

However, it is important to highlight that despite the advances of AI, the presence and importance of the human radiologist remains crucial, as AI should be considered a support and collaboration tool for healthcare professionals.

About The Author

Scroll to Top