4 Transformative technology Generative AI: More than a chatbot RADIOLOGY ‘Computer, why did the doctor take that MRI scan of my leg And what did it show ’: Popu- larized by OpenAI’s ChatGPT, generative artificial intelligence (AI) is already beginning to see practical applications in medical settings. The technology holds immense potential, with benefits for patients, clinicians, and even hospital administration, accord- ing to Shez Partovi, MD. We spoke with the Chief Innovation & Strategy Officer and Chief Business Leader of Enterprise Informatics at Philips about per- sonalized medical information, hospital workflow optimization, digital clinical twins, and other plans the company has for gener- ative AI in healthcare. Talking about how healthcare can benefit from this technology, Parto- vi does not seem afraid to set the bar high: ‘The shift that generative AI will bring is nothing like any- thing that we have seen. It will be like the discovery of penicillin or insulin; it will be the most tectonic shift for healthcare that we will see in our lifetime.’ Unlike predictive models, which rely on extensive datasets and meticulous training, generative AI finds associations be- tween data points to generate new content, the expert explains. To harness this potential, Philips is de- veloping several products based on the technology, some of which have already arrived in clinical practice. ‘In the future, people will talk to their medical record’ Among these is a tool that reph- rases physician’s notes in a way that is easier to understand. Pre- vious research has shown that medical information, even material intended for patients, is often too complex. ‘The idea is that the AI translates this complicated medical terminology to simple high-school level English and other languages,’ Partovi explains. With its ability to quickly generate customized text, generative AI will take this approach even further, Shez Partovi, MD Shez Partovi, MD is Chief Inno- vation & Strategy Officer and Chief Business Leader of Enterprise In- formatics at Royal Philips, where he drives cloud transformation and AI in healthcare. Previously, he spear- headed global business development efforts at Amazon Web Services. s p i l i h P © the expert predicts: ‘In the future, people will talk to their medical record. They can ask the AI why their doctor performed an MRI scan on them, or why they were prescribed specific medication, and it will explain the medical indi- cation in a way they can under- stand.’ With planned market intro- duction this conversational access to a patient’s medical record could be used to bridge the knowledge gap between doctors and patients, and help pa- tients make informed decisions about medical procedures, Partovi is convinced. this year, later A similar approach can be applied to provide physicians with relevant relevant background patient information, for example before consultations: ‘Reviewing a patient’s medical history can be very time-consuming. To mitigate this, the AI can quickly analyse the dataset and generate a summary which provides the physician with the in- formation.’ As a result, less time is spent on preparation, leaving more for the actual visit, Partovi argues. ‘More and more medical data is generated for each patient, which can prove overwhelming for the clinicians. For a three-minute visit, the doctor may have to spend up to 20 minutes to look through the medical records. We want to use AI to reverse that ratio, so the doctor has more time for the patient.’ Finally, generative AI can con- tribute to more efficient workflows – both on department and hospital levels, Partovi points out. For example, at the beginning of their shift, the director of a hospital’s emergency department could task the AI with generating a succinct overview of the situation – patient volume, staff attendance, even a prediction of upcoming workload changes – to be able to act and plan ahead. ‘This, too, will take place as a conversation with the AI, which represents a digital twin of the hospital,’ the expert describes, pointing out that the applications can benefit every member of the hospital team up to the adminis- trative level. Liquified data to enable comprehensive analyses However, to exploit this potential, the existing data must be liqui- dated, made available to the AI, he continues: ‘Almost every medical device is generating data, but most of this information is sitting inside silos, and the AI cannot make as- sociations because it is all separate. We have therefore created a sol- ution called Capsule, a platform to integrate all the medical device data. These can then be used by the AI for predictive and generative models.’ ■ Article: Wolfgang Behrends Deep Learning improves image quality, reduces radiation dose DL reconstruction in paediatric imaging Recent developments in deep learning techniques are enhanc- ing clinical imaging quality and reducing radiation exposure for patients while also maintaining diagnostic accuracy. The latest AI component to clinical imaging – referred to as deep learning re- construction (DLR) – is having a particular benefit in paediatric imaging, according to Dr Samuel Brady from Cincinnati Children’s Hospital Medical Center, US. Speaking during a series of webin- ars hosted by Canon Medical Sys- tems Europe, providing in-depth insights into aspects of paediatric imaging, his presentation specifi- cally focused on the impact of AI on CT image quality and dose. Brady discussed advantages and limitations of FBP (filtered back projection) and IR (iterative recon- struction) algorithms for CT image reconstruction. He noted that while IR is good at removing spurious image noise and helps see body structures/organs clearer, it can soften structural and organ bound- aries, leading to an image that may appear slightly blurry to the eye. However, DLR has demonstrated the ability to remove image noise while maintaining and sharp image and providing the ‘potential to further reduce radiation dose to patients undergoing CT,’ said Brady. As a result, the technique is increasingly preferred within pae- diatric imaging as it enables lower radiation dose, thus reducing risks for children where tissues are still developing and are more suscep- tible to radiation damage. Increasing diagnostic confidence The expert noted how CT recon- struction has undergone substan- tial changes, particularly with the ability of DLR to improve image quality and offer low-contrast de- tectability. Brady, who is Chief of the Clinical Medical Physics Sec- tion in the Medical Center’s Depart- ment of Radiology, said AI for CT reconstruction enables cleaner, noise reduced and sharper images with a stronger edge definition than current IR algorithms. ‘These improved images have been shown to increase diagnostic confidence for radiologist performance: they are able to see smaller, more subtle, structures with greater con- fidence – even at the lower radi- ation dose levels afforded by DLR.’ reduced The reduction of image noise – the statistical fluctuation in the recon- struction algorithm that overlays the image and can obscure the underlying anatomy – has been a major benefit, the expert pointed out. ‘For the last 15 years, IR al- gorithms have image noise in CT images, but at the cost of softening organ boundaries, giv- ing the image a blurry, soft, and sometimes plastic look. That inter- feres with radiologist’s ability to identify low-contrast objects, such as cancer or infections, in the image. DLR now removes the image noise while keeping object edges sharp.’ ‘A huge win for paediatric imaging’ Historically, the only way to re- move image noise in a CT image was to increase patient radiation exposure. But with DLR trained to differentiate between noise and anatomical structure, image noise can be removed without turning up the radiation dose, said Brady. He pointed to Canon’s proprietary DLR algorithm called “Advanced Intelligent Clear-IQ Engine (AiCE)” on CT platforms as advancing the field and particularly the new ver- sion of DLR, “Precise IQ Engine (PIQE)”, designed to further in- crease object-edge sharpness. ‘The ability to see small objects sharper and more conspicuous is a huge win for paediatric imaging, since children have smaller internal or- gans, vasculature, and other struc- tures,’ said Brady. He said DLR is making a difference by enabling low noise and low dose imaging in CT. For IR, the consensus was that dose reduction should be limited to 15–30% to preserve low-contrast detectability. But Brady pointed out that initial phantom and patient observer studies in DLR have shown accept- able dose reduction between 44–83% for both low- and high- contrast object detectability tasks. Ultrasound for abdominal emergencies A second element of the Canon webinar saw paediatric radiologist Dr Elisa Aguirre from Hospital 12 de Octubre in Madrid, Spain, out- line how abdominal symptoms are the most common reason for pae- diatric emergency department vi- sits, with ultrasonography usually the first line and main technique in Dr Samuel Brady Dr Samuel Brady is Chief of the Clinical Medical Physics Section in the Department of Radiology, Cincin- nati Children’s Hospital Medical Center and Associate Professor of Radiology, University of Cincinnati. As a medical physicist, his primary experience lies in developing and maintaining protocols for imaging ef- ficacy and safety with a particular in- terest in dose-minimization for CT imaging techniques. enabling the diagnosis. She ex- plained how Doppler ultrasound and microvascular imaging can add relevant information in paediatric abdominal emergencies. ■ Report: Mark Nicholls EUROPEAN HOSPITAL Vol 33 Issue 3/24