Clean data is key to advancing AI in medical imaging: Prakash KS, Siemens Healthineers

As Generative Artificial Intelligence (GenAI) reshapes industries, its impact on healthcare, particularly medical imaging, is gaining momentum. AI-driven automation is enhancing diagnostics, improving efficiency, and supporting clinicians in ways that were once unimaginable. However, challenges such as data quality, bias, and regulatory compliance remain critical concerns. 

Siemens Healthineers, a medical technology company, is integrating Gen AI to improve its imaging and diagnostic solutions. In a discussion with TechCircle, Prakash KS, Head of GenAI CoC at Siemens Healthineers Global Development Center, explained how the company is using Gen AI, tackling data biases, and shaping the future of AI in healthcare. Edited Excerpts:

How do you integrate Gen AI into your products, services, and medical imaging? 

Our GenAI strategy focuses on three key areas: internal tools, serviceability and training, and product innovation. Internally, we integrate Gen AI across various systems, including product lifecycle management (PLM), customer relationship management (CRM), supply chain management (SCM), and doctor scheduling. With Gen AI now commoditized, we leverage existing tools and services to enhance efficiency, streamline operations, and accelerate time-to-market. This transformation has significantly impacted the entire PLM cycle. 

In serviceability and training, we use GenAI to improve equipment maintenance through predictive and preventive analysis. By feeding collected data into small language models, we enhance diagnostics and equipment serviceability. Additionally, our parent company’s My Learning World platform employs Gen AI to train healthcare professionals more effectively, covering both medical knowledge and equipment usage. 

For product innovation, we are integrating Gen AI across our imaging value chain. We are prototyping AI-driven automation for report generation, disease detection, image segmentation, and historical data integration. Our Smart Imaging Value Chain applies AI to enhance detection, segmentation, and reporting while ensuring minimal workflow disruption for clinicians and hospitals. We are also exploring Gen AI for MRI image reconstruction and AI RAD companions. By systematically embedding GenAI into our products, services, and internal operations, we enhance efficiency while maintaining the continuity of existing workflows. 

How does using GenAI in clinical images and reports improve diagnostic accuracy and efficiency? 

In the early days of automation, imaging processes like MRI scans produced thousands of images, making manual analysis impractical. Today, algorithms can perform automatic detection, segmentation, and pattern matching. However, these analyses are not yet integrated with existing clinical commentary, which includes patient symptoms, disease history, and data from other diagnostic reports like ECGs and blood tests. Previously, this integration wasn’t possible, or it had to be done as a separate post-processing step. 

Now, large language models can analyze images, extract insights, cross-reference data from other reports, and feed it into AI models. This allows for deeper analysis and identification of previously unseen patterns. With automation, Gen AI can now combine multiple data sources and detect patterns with greater accuracy than ever before.

How does AI-powered image segmentation help radiologists detect and diagnose conditions more effectively? 

When you get a CT scan, a specialist reviews the images. Since their time is limited, they examine a sample of the most relevant images selected by the system. 

Now, technology allows profiling of disease states and organs, making predictions, but always with human oversight. The radiologist makes the final decision, with AI acting as an assistant. Instead of reviewing a few images and making a judgment, the radiologist receives an AI-generated summary highlighting potential disease areas or lesions. This directs their attention to key areas for further examination. 

The AI does not replace human expertise. If it is unsure, it will indicate that rather than providing incorrect results. This improves efficiency and accuracy, helping radiologists catch details they might otherwise miss. If the AI's findings match the radiologist’s, it serves as an additional validation. 

AI also eliminates errors due to fatigue or time constraints, as it processes large amounts of data without bias. Its role is to assist, ensuring thorough and accurate diagnoses.

How is GenAI redefining radiology? Does it complement radiologists or fundamentally change their roles? With concerns about AI replacing human expertise across industries, how do you see AI working alongside radiologists rather than replacing them? 

Last year, for the first time, a model outperformed a radiologist not only in expertise but also in selecting the right treatment plan based on factors like age and gender. This has raised concerns that AI could replace radiologists, but these fears are exaggerated, though not unfounded. 

Patients still prefer a human to review their diagnosis, even if AI provides an initial assessment. Human validation remains crucial. While AI can measure empathy-related factors, it cannot truly personalize care in the way a human radiologist can. AI can certainly automate repetitive tasks, but it cannot replace the core responsibilities of a radiologist. 

Radiologists spend 50 to 70 percent of their time on administrative tasks, such as transferring images, compiling reports, and reviewing prior information. AI can automate these processes, but human oversight will always be necessary. Even as AI integrates into workflows, its role will be to assist rather than replace, allowing radiologists to focus on complex cases that might otherwise be overlooked due to time constraints. 

Beyond radiology, AI is set to transform healthcare by making treatments more personalized. Just as online music streaming shifted music from a standardized public experience to a highly individualized one, AI will do the same for healthcare. Drug delivery, treatment plans, and radiology decisions will be tailored to individual patients, using historical data and cohort comparisons to adjust thresholds. 

This shift will be deeply transformative, with AI reshaping both healthcare and education. In healthcare, AI will drive more personalized care, improving patient outcomes while keeping human expertise at the center. 

We've discussed how AI is transforming medical imaging, but what are the biggest challenges in scaling its adoption? 

The biggest challenge is not cost or adoption but clean data. Medical workflows, instruments, and scanners generate vast amounts of data, but AI models require clean data to function effectively. The accuracy of AI outputs depends on the quality of the data fed into them. The current approach—cleaning up data after it has been collected—is inefficient. Instead, data should be structured and usable at the point of generation. This shift is already happening, but it remains a key challenge. 

Another challenge is integrating AI into existing medical workflows without disrupting clinicians, physicians, or radiologists. AI should work seamlessly with interoperable data, incorporating ECG scans, medical images, and lab reports into a unified workflow. While partial integration exists, full interoperability is necessary. 

Regulatory compliance is also critical. AI in healthcare must adhere to legal and security requirements while minimizing errors. Even a single mistake in a high-stakes environment can lead to mistrust and hinder adoption. To ensure reliability, AI models must deliver near-perfect accuracy. 

Key challenges include clean data, workflow integration, regulatory compliance, and interoperability, all while maintaining strong cybersecurity. 

How do you see the future of AI in healthcare beyond imaging? 

AI is set to make healthcare more personal and efficient. Drug discovery, drug delivery, and personalized treatment plans will become more integrated. AI will also improve operational efficiency, such as scheduling beds and managing lab workloads, leading to shorter wait times. Patients won’t have to spend an entire day waiting for basic health checks. 

Early detection will improve, with some conditions that currently require scans being detectable through simple blood tests. Clinical decision support is being developed across the industry, including by Siemens Healthineers, to enhance diagnosis and treatment. 

One key innovation is the concept of a digital twin. Siemens Healthineers is working on creating virtual models of patients using medical records, past treatments, scans, blood tests, and other health data. These models can help detect early-stage conditions that might otherwise go unnoticed. 

Digital twins will also enable more personalized treatments. Instead of patients questioning whether a prescribed drug is safe or if the dosage is appropriate, AI will help tailor treatments to individual needs. The future of healthcare is moving toward precision medicine, digital twinning, and highly personalized care. 

How does your company ensure that the data used to train GenAI models is clean, unbiased, and representative of a diverse patient population? What steps do you take to prevent data drift and inconsistencies in medical imaging datasets?

We invest time, effort, and money to collect data correctly. Our database contains about 1.2 billion medical records, including images, operational data, clinical data, and scans. We use this data as input while ensuring strict selection processes. We do not accept all incoming data; instead, we follow rigorous standards for scanning, storing, and transporting data for model training. These standards go beyond regulatory requirements to ensure data quality, lack of bias, and diversity. We check for regional, gender, and age-related biases to maintain a fair and representative dataset. 

We also invest in Federated Learning. With hospital and clinic permissions, we process data at the scanning location instead of transferring raw images or medical records. A smaller model extracts insights on-site, and only those insights are used to train our larger model. The original data remains at the source to prevent unnecessary data movement. 

Our approach to clean data involves three key steps. First, we work with a large dataset collected over years. Second, we ensure transparency and compliance at every stage of data collection, exceeding regulatory requirements because we handle sensitive medical information. Third, Federated Learning allows us to stay up-to-date without directly storing new patient data. Instead, we extract insights to improve our models.

Even after training, we continuously test the model for bias, accuracy, completeness, and edge cases. It undergoes rigorous verification, validation, and clinical testing before deployment. Given the high stakes, we take every measure to avoid errors. 

---