AI in Radiology Practices

AI just ran its own multidisciplinary tumor board. And nailed the diagnosis + treatment. This was a full-stack oncology reasoning engine—pulling from imaging, pathology, genomics, guidelines, and literature in real time. A new paper in Nature Cancer describes how researchers built a GPT-4-powered multitool agent that: • Interprets CT & MRI scans with MedSAM • Identifies KRAS, BRAF, MSI status from histology • Calculates tumor growth over time • Searches PubMed + OncoKB • And synthesizes everything into a cited, evidence-based treatment plan In short: it acts like a multidisciplinary team. Results : • Accuracy jumped from 30% (GPT-4 alone) to 87% • Correct treatment plans in 91% of complex cases • Every conclusion backed by a verifiable citation This is bigger than oncology. Any field that relies on multi-modal data and cross-domain reasoning—like my field of GI ( GI + Mental Health+ Nutrition + Excercise ) could benefit from this collaborative AI architecture. Despite the visual, it doesn’t replace the human team—it augments it. Providers still decide. But now, they do it faster, with more context, and less cognitive fatigue. #AI #HealthcareonLinkedin #Healthcare #Cancer

Case Tuesday: Lung Cancer Screening A 62-year-old lifelong smoker comes in for a routine low-dose CT as part of a lung cancer screening program. For the radiologist, these scans are among the most high-stakes and high-volume reads: Thousands of screening scans across a population Tiny nodules that may represent early-stage cancer The pressure to detect disease early without overwhelming patients and clinicians with false alarms The challenge: Huge workloads in screening program; subtle nodules can be missed, especially in noisy low-dose images; tracking growth over time requires precision and consistency This is where #AI can make a profound difference: Automatically detects and flags pulmonary nodules, even very small ones Measures and tracks nodule growth across serial scans Standardizes reports, supporting consistent follow-up recommendations The radiologist remains central, applying expertise and judgment. But AI provides a safety net that scales helping ensure no early cancer is overlooked, even in national screening programs. The impact: Earlier detection when lung cancer is most treatable, reduced false negatives and unnecessary anxiety for patients, greater efficiency in high-volume screening programs. I believe lung cancer screening is one of the clearest demonstrations of AI’s value: not just improving workflows, but saving lives on a population scale. How do you see AI enabling broader adoption of lung cancer screening especially in health systems where radiologist resources are stretched thin #CaseTuesday #LungCancerScreening #Radiology #AIinHealthcare #PopulationHealth #GEHealthcare

I am tremendously excited about the real-world impact of our latest publication on #AI #Biomarkers in Nature Medicine: https://lnkd.in/dv-7aS7Y Even in the US barely half of #lungcancer patients are tested for #EGFR mutations, for which targeted therapies readily exist. We have worked for many, many years now to try to overcome this gap with AI for H&E slides to offer patients a fast and cost-effective solution to get the right treatment. The point of this work is not only that we actually built it, but that Gabriele Campanella and Chad Vanderbilt organized a consortium and created the infrastructure for the first real-world, real-time deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection. 𝙋𝙧𝙤𝙨𝙥𝙚𝙘𝙩𝙞𝙫𝙚𝙡𝙮!   𝐌𝐞𝐞𝐭 𝐄𝐀𝐆𝐋𝐄 (EGFR AI Genomic Lung Evaluation): ✅ 𝟎.𝟖𝟗 𝐀𝐔𝐂 in a 𝐩𝐫𝐨𝐬𝐩𝐞𝐜𝐭𝐢𝐯𝐞 silent trial with clinical-grade performance. 🌍 Generalizes 𝐚𝐜𝐫𝐨𝐬𝐬 𝐡𝐨𝐬𝐩𝐢𝐭𝐚𝐥𝐬 𝐚𝐧𝐝 𝐜𝐨𝐧𝐭𝐢𝐧𝐞𝐧𝐭𝐬 with robustness and reproducibility. 🔬 Validated on 𝐢𝐧𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐜𝐨𝐡𝐨𝐫𝐭𝐬, 𝐦𝐮𝐥𝐭𝐢𝐩𝐥𝐞 𝐢𝐧𝐬𝐭𝐢𝐭𝐮𝐭𝐢𝐨𝐧𝐬, 𝐚𝐧𝐝 𝐬𝐜𝐚𝐧𝐧𝐞𝐫𝐬. 🧪 𝟒𝟑% 𝐫𝐞𝐝𝐮𝐜𝐭𝐢𝐨𝐧 𝐢𝐧 𝐫𝐚𝐩𝐢𝐝 𝐦𝐨𝐥𝐞𝐜𝐮𝐥𝐚𝐫 𝐭𝐞𝐬𝐭𝐬, preserving biopsy tissue for full genomic profiling. ⚡ 𝐃𝐞𝐥𝐢𝐯𝐞𝐫𝐬 𝐫𝐞𝐬𝐮𝐥𝐭𝐬 𝐢𝐧 𝐮𝐧𝐝𝐞𝐫 𝟏 𝐡𝐨𝐮𝐫, compared to 2–3 weeks for NGS. 🚀 A foundational step toward regulatory approval and 𝐀𝐈-𝐢𝐧𝐭𝐞𝐠𝐫𝐚𝐭𝐞𝐝 𝐜𝐥𝐢𝐧𝐢𝐜𝐚𝐥 𝐰𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬.   We have worked on Computational Biomarkers in Pathology continuously for over a decade starting with AI for predicting SPOP in prostate cancer from H&E in 2015, but seeing everything come to fruition at such a scale in 2025 is very humbling. AI, when done right, can give real, tangible help to cancer patients. 𝑰𝒕 𝒊𝒔 𝒐𝒖𝒓 𝒓𝒆𝒔𝒑𝒐𝒏𝒔𝒊𝒃𝒊𝒍𝒊𝒕𝒚 𝒕𝒐 𝒎𝒂𝒌𝒆 𝒊𝒕 𝒂 𝒓𝒆𝒂𝒍𝒊𝒕𝒚! I am deeply grateful to everyone on this most amazing team: Gabriele Campanella, Neeraj Kumar, Ph.D., Swaraj Nanda, Siddharth Singi, Eugene Fluder, Ricky Kwan, Silke Mühlstedt, Nicole  Pfarr, Peter Schüffler, Ida Häggström, Noora Neittaanmäki, Levent Akyürek, Alina Basnet, Tamara Jamaspishvili, Michel Nasr, Matthew Croken, Fred Hirsch, Arielle Elkrief, Helena Yu, Orly Ardon, Greg Goldgof, Meera Hameed, Jane Houldsworth, Maria E. Arcila, Chad Vanderbilt #AI #ComputationalPathology #Biomarkers #AIinHealthcare #DigitalPathology #PrecisionMedicine #LungCancer #EGFR #NatureMedicine #FoundationModels #EAGLEModel #EAGLE #Oncology

Duke researchers just simulated a nationwide lung cancer screening trial entirely in silico. Clinical trials like NLST shaped medical imaging but required 50,000+ patients and $256M over 8 years. Now, a Virtual Lung Screening Trial (VLST) shows how we might do this in weeks instead of decades. 𝗩𝗟𝗦𝗧 𝗶𝘀 𝗮𝗻 𝗲𝗻𝗱-𝘁𝗼-𝗲𝗻𝗱 𝘃𝗶𝗿𝘁𝘂𝗮𝗹 𝗶𝗺𝗮𝗴𝗶𝗻𝗴 𝘁𝗿𝗶𝗮𝗹 𝘁𝗼 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝗖𝗧 𝘃𝘀. 𝗖𝗫𝗥 𝗳𝗼𝗿 𝗹𝘂𝗻𝗴 𝗰𝗮𝗻𝗰𝗲𝗿 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻, 𝗱𝗼𝘄𝗻 𝘁𝗼 𝘀𝗰𝗮𝗻𝗻𝗲𝗿, 𝗿𝗲𝗮𝗱𝗲𝗿, 𝗮𝗻𝗱 𝗽𝗮𝘁𝗶𝗲𝗻𝘁 𝗹𝗲𝘃𝗲𝗹𝘀.  1. Simulated 294 virtual patients with and without nodules, using XCAT phantoms with diverse age, sex, BMI, and race distributions.  2. Built scanner-specific CXR and CT simulations using DukeSim, matched to real-world image quality from NLST-era hardware.  3. Developed virtual readers using 2D and 3D RetinaNet models trained on public datasets (LUNA16, NODE21), interpreting all patient scans.  4. Achieved CT AUC of 0.92 (95% CI: 0.89–0.95), significantly outperforming CXR at 0.72 (95% CI: 0.67–0.77) in patient-level cancer recall. Couple thoughts:  • Could fine‑tuning a lightweight transformer‑based detector (e.g., Detection Transformer) to potentially boost sensitivity to small or irregular nodules?  • Adding longitudinal simulation of lesion growth and incorporating a survival‑prediction head could expand the trial from a recall task to prognostic modeling  • Using publicly available XCAT phantoms and DukeSim’s validated ray‑tracing/Monté Carlo framework is great for reproducibility and for community adoption. I hope it gets open-sourced! Unless it already is and I just can't find it haha Here's the awesome work: https://lnkd.in/gDmszB9N Congrats to Fakrul Islam Tushar, Joseph Lo, Ehsan Samei and co! I post my takes on the latest developments in health AI – 𝗰𝗼𝗻𝗻𝗲𝗰𝘁 𝘄𝗶𝘁𝗵 𝗺𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆 𝘂𝗽𝗱𝗮𝘁𝗲𝗱! Also, check out my health AI blog here: https://lnkd.in/g3nrQFxW

Clinical decision-making in oncology is complex, requiring the integration of multimodal data and multidomain expertise. We developed and evaluated an autonomous clinical artificial intelligence (AI) agent leveraging GPT-4 with multimodal precision oncology tools to support personalized clinical decision-making. The system incorporates vision transformers for detecting microsatellite instability and KRAS and BRAF mutations from histopathology slides, MedSAM for radiological image segmentation and web-based search tools such as OncoKB, PubMed and Google. Evaluated on 20 realistic multimodal patient cases, the AI agent autonomously used appropriate tools with 87.5% accuracy, reached correct clinical conclusions in 91.0% of cases and accurately cited relevant oncology guidelines 75.5% of the time. Compared to GPT-4 alone, the integrated AI agent drastically improved decision-making accuracy from 30.3% to 87.2%. These findings demonstrate that integrating language models with precision oncology and search tools substantially enhances clinical accuracy, establishing a robust foundation for deploying AI-driven personalized oncology support systems. Paper and research by Dyke Ferber,  Jakob Nikolas Kather and larger team. The excerpt above is from the author's abstract

This morning the results of our randomised trial on the impact of AI triage of primary care chest X-ray in England were published https://lnkd.in/euMAB6MW LungIMPACT is one of the largest radiology RCTs outside of screening; 93,326 CXR from 5 NHS hospital sites randomised. This is not a before-after study, there were no pathway changes, no additional reporting capacity, no ring fenced scanning slots. Co-primary outcomes, time to diagnosis of lung cancer and time to chest CT. We know early, rapid and accurate diagnosis improves outcomes for patients, and the strength of our trial design combined with the potential patient benefits is why we were one of Nature Medicine's "11 trials to shape medicine in 2024" The results? · 558 lung cancers diagnosed, 13,347 chest CT scans performed in 86,945 patients, between 6-23 months follow up & linked data across cancer registries. · AI triage & worklist prioritisation significantly reduced the time to CXR report, from 47 to 34 hours (p<0.001) · AI triage & prioritisation does not significantly shorten the time to CT (53 days both arms, p=0.31) or the time to lung cancer diagnosis (44 days prioritised vs 46 days without, p=0.84) . These results held across all sites, on all sensitivity analyses. · No differences in time to urgent lung cancer referral, time to treatment start, stage of lung cancer. · Nearly 30% of chest X-rays had a discordance between AI output and report - our expert team reviewed 28,261 CXRs, and found 6,750 actionable findings Strengths? The robust control, effective randomisation, diverse and dispersed NHS sites across geographies, volumes and populations. No other confounding changes, the pathway stayed the same, we quantified the contribution that AI prioritisation could have. Our experts reviewed a staggering amount of discordant chest X-rays, almost one-third of the total. Limitations?  A single AI vendor, all chest X-rays having AI decision support so comparison without AI is not possible Conclusion? AI prioritisation of chest X-rays referred from UK primary care has no impact on the lung cancer pathway. That most lung cancers were flagged as abnormal by AI means that these results are vendor neutral, prioritisation without additional resourced pathway changes do not work. AI deployments should not include worklist prioritisation in this setting. The takeaway? Alone, AI does not improve the outcomes for lung cancer patients in the NHS. Policy makers need to understand that patients and pathways rarely respond in a linear way. LungIMPACT was a true team effort, led by David Baldwin with Lesley Smith Janette Rawlinson Arjun Nair Iain Au-Yong Bindu George Madava Djearaman Richard Lee Neal Navani Siyabonga Ndwandwe, AFHEA Caroline Clarke Andy Creeden Josh Newsome Sylvia Abakporo Richard Tucker Indrajeet Das James Hathorn

The age of AI has truly begun, and its impact will surpass the internet, mobile, and cloud revolutions combined.   In 2022, DeepMind released AlphaFold. Using AI, they predicted the structure of nearly 200 million proteins. To put that in perspective until recently, discovering the structure of one protein could take the equivalent of a PhD’s lifetime of work. Why does this matter? Because most diseases are driven by proteins. Understanding how proteins fold and how other molecules interact with them fundamentally transforms how we discover drugs and treat disease. I genuinely believe that with the help of AI, we will find solutions to challenges like cancer and Alzheimer’s that have haunted humanity for decades. We have seen this transformation firsthand. Tuberculosis remains one of the biggest health challenges in South Asia, accounting for nearly 25 percent of global TB cases. In 2016 to 17, we spun out Qure.ai with a simple but ambitious goal to use AI to make early diagnosis accessible at scale. Today, Qure.ai can analyze X rays for tuberculosis and lung cancer in seconds, at one dollar per scan. For context, even with the best tools available, radiologists can miss findings nearly 25 percent of the time. In the Philippines, mobile X ray vans were sent to remote villages. Earlier, scans had to be sent back to cities for diagnosis often taking six weeks. By then, patients might infect others or never return for results. Now, with AI embedded directly into the machines, diagnosis happens instantly. Patients are told immediately whether they need specialist care, medication, or simply a follow up later. To date, Qure.ai has impacted over 25 million lives diagnosing TB, lung cancer, and paving the way for many more conditions in the future. This is what AI makes possible when applied with purpose. Not efficiency alone but scale, equity, and human impact. We are only at the beginning.

𝐀𝐈 𝐀𝐠𝐞𝐧𝐭 𝐀𝐜𝐡𝐢𝐞𝐯𝐞𝐬 𝟗𝟏% 𝐀𝐜𝐜𝐮𝐫𝐚𝐜𝐲 𝐢𝐧 𝐂𝐚𝐧𝐜𝐞𝐫 𝐃𝐢𝐚𝐠𝐧𝐨𝐬𝐢𝐬 Oncology decision-making is notoriously complex. Clinicians must integrate histopathology images, radiology scans, genetic profiles, and ever-evolving treatment guidelines to make personalized care decisions. It's a cognitive challenge that even experienced specialists find demanding. A new study by Ferber et al. in Nature Cancer shows how an autonomous AI agent tackled this complexity head-on—and the results are striking. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: Current AI approaches in healthcare often work in isolation—analyzing single data types or providing generic responses. But real clinical decisions require synthesizing multiple sources of evidence simultaneously, something that has remained challenging for AI systems. 𝗞𝗲𝘆 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻𝘀: ◦ 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐭𝐨𝐨𝐥 𝐢𝐧𝐭𝐞𝐠𝐫𝐚𝐭𝐢𝐨𝐧: Vision transformers detect genetic mutations directly from tissue slides, MedSAM segments tumors in radiology images, and the system queries precision oncology databases autonomously ◦ 𝐒𝐞𝐪𝐮𝐞𝐧𝐭𝐢𝐚𝐥 𝐫𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠: The agent chains tools together—first measuring tumor growth from imaging, then checking mutation databases, then searching recent literature ◦ 𝐄𝐯𝐢𝐝𝐞𝐧𝐜𝐞-𝐛𝐚𝐬𝐞𝐝 𝐜𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬: 75.5% accuracy in citing relevant medical guidelines, addressing the critical problem of AI hallucinations in healthcare 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: When tested on 20 realistic patient cases, the integrated system achieved 91% accuracy in clinical conclusions. Perhaps more telling: GPT-4 alone managed only 30% accuracy on the same cases—nearly a 3x improvement through tool integration. The agent successfully used appropriate diagnostic tools 87.5% of the time and provided helpful responses to 94% of clinical questions. 𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗶𝗰𝘁𝘂𝗿𝗲: This isn't about replacing oncologists—it's about augmenting clinical reasoning with AI that can process multiple data streams simultaneously. The modular approach means individual tools can be updated, validated, and regulated independently. While challenges remain around data privacy and regulatory approval, this research points toward a future where AI agents serve as sophisticated clinical reasoning partners, helping doctors navigate the increasing complexity of modern medicine. https://lnkd.in/e52xBZj9 #AIinHealthcare #PrecisionOncology #ClinicalAI #DigitalHealth #MachineLearning #Oncology

AI won't replace oncologists, but it may help them outsmart cancer. Harvard's Sybil model can flag early lung-cancer risk on CT scans months before visible signs appear. Other AI systems now match genomic profiles to targeted drugs, accelerating personalized treatment planning. As Dr. Marc Siegel noted, AI's greatest promise lies in early detection and precision medicine, helping doctors act before cancer spreads, not after. The breakthroughs are already happening across radiology and oncology: AI that spots what humans miss, triages risk, and personalizes therapy. AI in healthcare shouldn't replace clinical expertise, it should amplify it. Giving physicians superhuman pattern recognition while they keep the human judgment and compassion. https://lnkd.in/eHCGPb26

Artificial intelligence improves ling cancer diagnosis. New AI-based digital platform enables extremely fast and accurate analysis of tissue sections from lung cancer patients. Germany. August 23 2024. Excerpt: A team of researchers from the University of Cologne’s Faculty of Medicine and University Hospital Cologne, led by Dr Yuri Tolkach and Professor Dr Reinhard Büttner, has created a digital pathology platform based on artificial intelligence. Note: The platform uses new algorithms developed by the team and enables fully automated analysis of tissue sections from lung cancer patients. The platform makes it possible to analyse digitized tissue samples on the computer for lung tumors more quickly and accurately than before. The study. ‘Next generation lung cancer pathology: development and validation of diagnostic and prognostic algorithms’ has been published in the journal Cell Reports Medicine. Lung cancer is one of the most common cancers in humans and is associated with a very high mortality rate. Today, the choice of treatment for patients with lung cancer is determined by pathological examination. Pathologists can also identify molecularly specific genetic changes that allow for personalized therapy. Over the past few years, pathology has undergone digital transformation. Microscopes are no longer needed. Typical tissue sections are digitized and then analyzed on a computer screen. Digitalization is crucial for the application of advanced analytical methods based on artificial intelligence. By using artificial intelligence, additional information about the cancer can be extracted from pathological tissue sections – that would not be possible without AI technology. The published study also conveys how the platform could be used to develop new clinical tools. The new tools not only improve the quality of diagnosis, but also provide new information about the patient’s disease, such as how the patient is responding to treatment,” explained physician Dr Yuri Tolkach from the Institute of General Pathology and Pathological Anatomy at University Hospital Cologne, who led the study. In order to prove the broad applicability of the platform, the research team will conduct a validation study together with five pathological institutes in Germany, Austria and Japan. Publication: https://lnkd.in/eyCGaQjJ Direct link to the publication is also available in enclosed announcement. https://lnkd.in/e_Uhq72b