How AI is Moving Upstream in Radiology

Over the past decade, advances in imaging hardware have dramatically increased the volume and complexity of radiological data. MRI and CT scans now generate thousands of images per study, while population-scale screening programmes have increased radiologists' workload. However, this growth is stunted by persistent workforce constraints, especially in high-volume healthcare systems such as India’s where demand for imaging has outpaced the availability of trained specialists.

According to a study published in the International Journal of Health Technology and Innovation, India has only 20,000 radiologists serving 1.4 billion people.

From a research perspective, this imbalance between demand and available expertise is now a central reason why AI is becoming critical to radiology. 

“The big issue is that human expertise is lacking globally,” Tobias Heimann, Head of AI Research at medical technology company Siemens Healthineers, tells AIM. “With the new technology, we have the opportunity to maintain and extend a high level of healthcare around the globe despite ageing populations, more chronic disease and shortages of skilled workers.”

From Workflows to Image Acquisition

Early uses of AI in radiology largely focused on post-processing, including detecting anomalies, flagging suspicious findings, and assisting with reporting and workflows. What is changing now is the point at which AI enters the imaging pipeline.

At Siemens Healthineers, AI is being integrated across the entire imaging value chain—from acquisition and reconstruction to interpretation. “AI has moved from experimental use into routine clinical imaging,” says Kalavathi GV, Executive Director and Head of the Global Development Center at Siemens Healthineers. The shift, she notes, has been driven by rising patient volumes, workforce constraints, and the need for faster, consistent diagnostics.

AI’s most tangible impact can be seen in research, especially when it comes to AI-based image reconstruction in MRI.“Depending on the sequence, you can reduce scan times dramatically or double the resolution within the same time. It really shifted the boundaries of what you could do with an MRI machine,” Heimann says.

In MRI, AI-based reconstruction techniques enable high-quality images to be produced from compressed raw data. Siemens’ Deep Resolve technology, for example, uses trained neural networks to reduce noise, improve spatial resolution, and significantly shorten scan times. Shorter scans have multiple downstream effects: reduced motion artefacts, improved patient comfort, and better utilisation of expensive scanner infrastructure.

AI-driven reconstruction is also enabling new hardware approaches. Heimann highlights the role of AI in making low-field MRI systems clinically viable. “With smaller, lighter, and more affordable scanners, the image quality would traditionally not be competitive,” he explains. “But because we are using AI for reconstruction, the image quality we can get out of these systems is actually comparable to standard scanners.”

CT imaging is seeing similar changes. AI-based applications now automate tasks such as coronary calcium detection, organ segmentation, and volume calculation—functions that previously required manual effort. According to Siemens Healthineers, these capabilities support workflows in cardiovascular care, oncology, neurology, and emergency medicine, while also enabling lower radiation doses and faster exams, including on newer photon-counting CT systems.

Reducing Scan Time, Improving Throughput

For patients, scan duration is not an abstract metric. MRI procedures can be physically uncomfortable and anxiety-inducing, particularly for cardiac patients or those unable to remain still for long periods. For clinicians and hospital administrators, scan time directly affects throughput and access to care.

Philips has focused much of its AI development on reducing acquisition time while maintaining diagnostic quality. At its Innovation Campus in India, engineers work on software that powers Philips’ global MRI systems. According to Philips executives, AI-enabled workflows can reduce MRI scan times from several minutes to under a minute.

The implications extend beyond patient comfort. Faster scans enable more patients to be imaged on the same system, improving access in settings where imaging infrastructure is limited. “Not everybody can hire a trained cardiologist or radiologist, and even buying an MRI machine is expensive,” Arvind Vaishnav, VP, Head of Philips Innovation Campus, tells AIM, pointing to the need for technology that allows clinicians to do more with existing resources.

From a research standpoint, reducing scan time is one of the most meaningful patient-facing outcomes of AI adoption. Heimann notes that these gains are already in clinical use. “This AI-based reconstruction is in place,” he says. “It allows you to reduce the time the patient needs to spend inside the machine, depending on the body region and the sequence.”

AI is also standardising and simplifying scanner operation. Automated patient positioning, protocol selection, and calibration reduce variability between operators and lower the skill threshold required to perform complex imaging studies. In CT, intelligent guidance systems assist with scan planning, dose optimisation, and error prevention, helping ensure consistent image quality across sites.

Decision Support, Not Decision-Making

One of the most persistent concerns is automation replacing clinical judgement. Equipment manufacturers and researchers are careful to frame current systems as decision-support tools rather than autonomous diagnosticians.

At Siemens Healthineers, AI is positioned to offload repetitive and time-consuming tasks—measurements, segmentation, and data processing—so that radiologists can focus on complex cases. “The radiologist is accountable for the final decision, interpretation, and reporting,” says Kalavathi.

Heimann reinforces this distinction from a research perspective. While AI can reliably detect specific findings, full diagnostic autonomy remains out of its scope. “Most AI solutions today focus on individual findings,” he explains. “They might detect lung nodules or specific cardiac markers, but the radiologist still needs to check the findings, approve them, and integrate them into a report. That human oversight is essential.”

In mammography, AI has shown particular promise in high-volume screening settings. According to a 2025 study on The Lancet, AI can reduce the number of human readers required. “In breast cancer screening, AI can support single-reader workflows instead of two radiologists per image,” he says. “That effectively doubles screening capacity, which makes a significant difference when workforce availability is limited.”

Data Quality, Explainability and Regulation

Beyond performance metrics, AI’s deployment in radiology raises questions around data quality, explainability, and regulatory compliance. According to Siemens Healthineers, the primary challenge is not cost or adoption, but data integrity. “AI models are only as reliable or effective as the data that they are trained on,” Kalavathi notes.

From a research and regulatory standpoint, Heimann emphasises the importance of representative, global datasets. “If you want to deploy AI globally, you need to train it on global data,” he says, noting that population-level differences can influence model performance. At the same time, regulatory frameworks in many regions require local data validation.

Data privacy is also a central concern. Heimann explains that AI training uses de-identified datasets. “All personal information—patient names, IDs—is stripped before the data is transferred for training,” he says. “There is no way to re-identify the data we use.”

Explainability is another area of focus. Rather than producing opaque outputs, imaging AI systems increasingly provide visual overlays, quantitative measurements, and confidence indicators that allow radiologists to understand the system’s conclusions.

Standards-based interoperability also plays a role. Both Philips and Siemens emphasise adherence to key healthcare interoperability standards like DICOM and FHIR, enabling AI tools to integrate with existing hospital IT systems. From a regulatory perspective, compliance with frameworks such as the General Data Protection Regulation (GDPR) governs how patient data is processed and used in research and development, even outside Europe.

The India Lens

India occupies a distinctive position in the global radiology AI ecosystem. According to market research firm GlobalData, India accounts for 20% of the Asia-Pacific diagnostic imaging device market, representing a major centre for R&D and manufacturing.

Philips’ India Innovation Campus contributes software and design work for imaging systems deployed worldwide. Components such as MRI coils manufactured in India are used globally.

From a clinical and operational perspective, India’s imaging volumes make it an early testbed for workflow automation. Heimann notes that the lack of trained operators is a recurring challenge globally. This is driving research into more autonomous scanning, where systems automatically adjust parameters based on clinical intent.

What Changes and What Doesn’t

Despite rapid technical progress, the fundamentals of radiology remain unchanged. 

What is changing is the structure of radiology work. As AI takes on acquisition optimisation, reconstruction, prioritisation, and documentation, radiologists are being asked to engage more deeply with complex cases, multidisciplinary discussions, and patient interaction. 

Heimann sees the next phase as end-to-end workflow support. “The real efficiency gain will come when systems don’t just find abnormalities but draft structured reports that radiologists can review and approve,” he says.