AI in Radiology Practices

Today, Radiology published our latest study on breast cancer. This work, led by Felipe Oviedo Perhavec from Microsoft’s AI for Good Lab and Savannah Partridge (UW/Fred Hutch) in collaboration with researchers from Fred Hutch , University of Washington, University of Kaiserslautern-Landau, and the Technical University of Berlin, explores how AI can improve the accuracy and trustworthiness of breast cancer screening. We focused on a key challenge: MRI is an incredibly sensitive screening tool, especially for high-risk women—but it generates far too many false positives, leading to anxiety, unnecessary procedures, and higher costs. Our model, FCDD, takes a different approach. Rather than trying to learn what cancer looks like, it learns what normal looks like and flags what doesn’t. In a dataset of over 9,700 breast MRI exams—including real-world screening scenarios—our model: Doubled the positive predictive value vs. traditional models Reduced false positives by 25% Matched radiologists’ annotations with 92% accuracy Generalized well across multiple institutions without retraining What’s more, the model produces visual heatmaps that help radiologists see and understand why something was flagged—supporting trust, transparency, and adoption. We’ve made the code and methodology open to the research community. You can read the full paper in Radiology https://lnkd.in/gc82kXPN AI won't replace radiologists—but it can sharpen their tools, reduce false alarms, and help save lives.

Case Tuesday: Cardiac CT A patient presents with chest pain. The question is urgent: is this a heart attack waiting to happen, or something else? A CT coronary angiogram is performed. For the radiologist, this means carefully assessing coronary arteries, looking for stenosis, calcifications, and subtle plaques. The challenge: Coronary CTs generate hundreds of slices, often complex to interpret. Subtle plaques can be easily overlooked. Quantifying calcium scores and stenosis consistently takes significant time. This is where #AI is showing real promise: Automated calcium scoring to assess cardiovascular risk Plaque detection and quantification to support precise diagnosis Tools that standardize reporting and improve communication with cardiologists The radiologist’s expertise is essential in interpretation and clinical context but AI ensures that the assessment is faster, more reproducible, and more actionable. The impact: Earlier detection of coronary artery disease. Better risk stratification for patients with chest pain. Closer collaboration between radiology and cardiology teams As Chief Medical Officer at GE HealthCare, I see cardiac CT as a shining example of how AI doesn’t just enhance workflows it helps us move toward preventive, precision medicine that saves lives before catastrophe strikes. Do you see AI as the tipping point that will make cardiac CT more widely adopted as a first-line test for chest pain? #CaseTuesday #CardiacCT #AIinHealthcare #Radiology #HeartHealth #GEHealthcare

🩻🧠 GE HealthCare #Wins FDA #510(k) for #True #Definition #DL — #Deep #Learning #CT Pushes High-Resolution Lung, Bone & Inner Ear Imaging to the Next Level A meaningful AI imaging infrastructure upgrade in radiology. Chad Rowland, Executive Director, Global Premium CT and Photon Counting, GE HealthCare, and @Dr. @Stefanie Bitschnau, Radiologist, RadioMed Corporation, highlighted why this matters: diagnostic confidence starts with the ability to see subtle anatomy clearly and fast. With FDA clearance, True Definition DL expands GE HealthCare’s deep learning CT stack beyond TrueFidelity DL + True Enhance DL, bringing: ✅ Higher spatial resolution in bone + lung ✅ Artifact suppression via dedicated neural network ✅ 1024 matrix high-resolution display ✅ Chest imaging in under 1 second ✅ Better visibility of small airways, nodules, trabecular bone, ossicles Why this matters 👇 1️⃣ 🫁 Earlier disease detection advantage The biggest clinical impact may be in: 🔹 interstitial lung disease 🔹 small airway disease 🔹 subtle pulmonary nodules 🔹 micro-fractures 🔹 inner ear erosions These are exactly the findings where improved spatial resolution changes patient pathways. 2️⃣ ⚡ AI solving the old CT trade-off Historically, sharper CT meant choosing between: higher dose ❌ slower scans ❌ limited coverage ❌ Deep learning reconstruction changes that equation by delivering detail without the traditional penalty. 3️⃣ 🏥 Practical scale beats premium hardware alone Photon counting CT gets the headlines, but scalable DL upgrades on installed systems may have broader real-world impact across hospitals and imaging centers. This is where AI becomes a true fleet multiplier. 4️⃣ 📈 Imaging economics Faster chest scans + better first-pass quality can reduce: 🔹 repeat scans 🔹 patient backlog 🔹 workflow bottlenecks 🔹 radiologist uncertainty That’s direct ROI for health systems. 💡 My takeaway This is another sign that AI in radiology is shifting from interpretation support to image formation itself. The future competitive edge in CT may increasingly come from software-defined image quality, where deep learning reconstruction upgrades extend the value of installed hardware fleets. That is a very scalable and sticky business model. #Radiology #CT #MedicalImaging #AI #DeepLearning #GEHealthCare #FDA510k #DigitalHealth #Diagnostics https://lnkd.in/gqqDvdex

Radiology AI has crossed an important line. It has moved from detecting one finding to generating a draft report for the entire examination. Most first-generation radiology AI products are narrow, task-specific models. One model is built for one job. FDA-cleared algorithms that flag large-vessel occlusion, pulmonary embolism, or pneumothorax are good examples. They have defined inputs and outputs, and a validation study can be designed around that single task. That is the model behind most FDA-cleared radiology AI to date. The newer generation, built on vision-language and foundation-model architectures, can evaluate much broader portions of an examination and generate a full draft report in plain language, not just a single flag. The verification task changes with it. With a narrow model, you know which specific output needs to be checked. A model drafting an entire report can be wrong across the examination while still sounding completely fluent. Confirming that it is right may require covering essentially the same ground the model covered. Northwestern Medicine is a useful place to look at what this means in practice rather than in theory. They deployed its own in-house generative model across an 11-hospital network and evaluated nearly 24,000 radiograph reports over five months. The average efficiency improvement was 15.5 percent, with some radiologists seeing gains of up to 40 percent, without compromising accuracy. The system drafts reports in each radiologist’s own dictation style. The Northwestern team also describes it identifying findings such as pneumothorax before the radiologist has opened the examination. The results were published in JAMA Network Open. https://lnkd.in/gC2YtcYy The question for us is no longer whether to pay attention. It is whether our organizations are building the knowledge needed to evaluate these systems and eventually deploy them on our own terms, before those decisions are made without enough radiology input. Send me a note. I would like to hear from radiologists who are leaning into this and from those who remain skeptical. For anyone who wants to go deeper, these two prior articles explore what foundation models could mean for radiology’s scope and why this may still be the best time in history to be a radiologist. The End of Diagnostic Silos. What Foundation Models Mean for Radiology’s Scope https://lnkd.in/gDVfCzeS The Best Time in History to Be a Radiologist https://lnkd.in/gX_6P5Ri

📌 Open-Source Medical Imaging AI Models (2024–2025) This curated list highlights the latest open-source AI models transforming medical imaging, from generalist vision-language foundations to specialized tools for segmentation, diagnosis, and report generation. Explore models across radiology, oncology, and multimodal analysis. Full links and details below. 👇 📌 Foundation & Multimodal Models • Rad-DINO – Self-supervised ViT trained on 1M+ chest X-rays • RayDINO – Large-scale DINO-based transformer for multi-task chest X-ray learning • Med-Gemini – Gemini-based model fine-tuned for multi-task chest X-ray applications • Merlin – Large 3D vision–language model for CT interpretation and reporting • RadFound – Radiology-wide VLM for report generation and question answering • LLaVA-Rad – Vision–language model for chest X-ray finding generation 📌 Segmentation Models • MedSAM2 – Promptable 3D segmentation model extending Segment Anything to medical imaging • FluoroSAM – SAM variant trained from scratch on synthetic X-ray/fluoro images • ONCOPILOT – Interactive model for CT-based 3D tumor segmentation in oncology 📌 Task-Specific / Tuned Models • MAIRA-2 – Enhanced CXR report generator with finding localization • CheXagent – Instruction-tuned multimodal model for chest X-ray tasks • RadVLM – Dialogue assistant for chest X-ray interpretation and reporting • Mammo-CLIP – CLIP-based model for mammogram classification and BI-RADS prediction • CheXFound – ViT model using GLoRI architecture for disease localization in X-rays Know a model that got missed? Drop it in the comments, let’s build this resource list together. 🤔 _________________________________________________ #ai #imaging #radiology #oncology #machinelearning

Could AI drafts—even imperfect ones—be a time-saver for radiologists when interpreting CT scans? Our pilot study using simulated AI reports found a 24% faster workflow, with accuracy intact. Q: What makes this study's approach unique? A: Instead of building an AI system, we used GPT-4 to simulate what AI-generated draft reports might look like. We deliberately introduced 1-3 errors in half the drafts to study how radiologists would handle imperfect AI assistance - a "Wizard of Oz" approach to prototype the future workflow. Q: How was the simulation study structured? A: We conducted a 3-reader crossover study with 20 chest CT cases. Each case was read twice: once with standard templates, and once with our simulated AI drafts. This controlled design let us directly compare the workflows. Q: What efficiency gains did you see with the simulated drafts? A: Median reporting time dropped from 573 to 435 seconds (p=0.003) - a 24% reduction. Two readers showed major improvements (717→398s and 361→322s), while one showed an increase (947→1015s). Q: Did the intentionally flawed drafts impact accuracy? A: Surprisingly, even with deliberately introduced errors in half the simulated drafts, the AI-assisted workflow showed slightly fewer clinically significant errors (0.27±0.52) compared to standard workflow (0.38±0.78). While not statistically significant, this suggests radiologists maintained their vigilance even with imperfect drafts. Q: How did radiologists respond to working with these simulated drafts? A: All 3 readers found the prototype system easy to use and well-integrated into their workflow. Two reported somewhat less mental effort, while one reported significantly reduced effort. Their likelihood to recommend it varied (scores of 5, 9, and 10 out of 10). Q: What's next? A: While these simulation results are encouraging, these are small scale pilot studies setting the stage for deeper validation. Link to short paper: https://lnkd.in/d-4aTJ69 Congratulations to stellar team of Julián Nicolás Acosta, Siddhant Dogra, Subathra Adithan, Kay Wu, MD 💫, Michael Moritz, Stephen Kwak

AI helps spot breast cancer on MRI scans by overlaying clear visual indicators, like colored heatmaps, that make suspicious areas stand out from healthy tissue. 🔬 The Core Visualization: Heatmaps & "Glowing" Cancer When AI analyzes a breast MRI, it creates a visual overlay to highlight areas of concern for the radiologist. This is primarily done through: · Spatially Resolved Heatmaps: The AI generates color-coded maps over the MRI image. Areas the model believes are abnormal are highlighted in color, allowing radiologists to focus on specific regions for further investigation. · Enhanced Delineation: Using advanced techniques like synthetic correlated diffusion imaging (CDI), AI can be optimized to make cancerous tissue appear to "glow" or "light up" next to healthy tissue, significantly improving the clarity of tumor boundaries. 🧠 How AI Achieves This Detection This capability is powered by AI models, often based on deep learning, that are trained on vast datasets of breast MRI exams. A key advancement is the development of explainable AI anomaly detection models. Unlike older systems: · They learn a robust pattern of what normal, benign breast tissue looks like. · They then flag significant deviations from this pattern, which helps in identifying malignancies even when such cases are rare in the training data. · This approach is particularly promising for screening populations, where cancer prevalence is low. 📊 Documented Performance and Potential Impact Research shows these AI systems are becoming highly effective clinical tools. The key findings are summarized below. Key Performance Metrics: · Detection Accuracy: Matches the performance of board-certified breast radiologists in clinical studies. · False Positive Reduction: Can potentially help avoid up to 20% of unnecessary biopsies in certain patient groups. · Generalizability: Models trained on large datasets (e.g., over 21,000 MRI exams) have shown robust performance when tested on external patient data from other institutions and countries. Primary Clinical Goal: The technology is designed as a powerful assistant to radiologists, not a replacement. It aims to improve reading efficiency, reduce unnecessary procedures, and help detect cancers that might otherwise be missed.

Think AI will replace radiologists? The truth is much weirder—and way more interesting... Radiology is about to go through an identity crisis. Because for over 100 years, medical imaging has been designed for one thing: the human eye. High-res, high-contrast, just the right number of slices—like a very expensive Instagram filter for tumors. But AI doesn’t care what the picture looks like. It doesn’t “see.” It detects. It can pull signal from noise, find disease in raw data, or predict cancer risk in a breast that looks totally normal to a human. So here’s the real shift: We’ve been building imaging machines for people to look at. Now we need totally different machines optimized for AI to think with. And that’s already happening. • Jonathan Rothberg’s Hyperfine | AI-Powered Portable MRI is so low-powered it’s unreadable by humans—but AI interprets it just fine. • Nanox Vision is making cloud-connected X-ray machines meant for AI-only triage at scale. • Qure.ai, founded by Prashant Warier, is screening for TB across rural India using chest X-rays and AI—with no radiologist in sight. • Lunit Cancer Screening, led by Brandon B. Suh, is getting FDA clearance for autonomous cancer detection and running large-scale real-world deployments. • And at MIT, Regina Barzilay’s Mirai model predicts breast cancer risk five years out—before anything shows up on a scan. So what happens to the radiologist? They don’t vanish. They evolve. Into a role that’s less about manually scanning slices and more about: • Synthesizing AI insights with clinical judgment • Validating outputs across multiple data sources and modalities • Guiding diagnostic strategy when the answer isn’t obvious—and the stakes are high • Overseeing the safety, bias, and reliability of AI tools in real-world care They may never even look at the image. There may not even be a meaningful image to look at (much like QR codes today) Because the image isn’t the diagnosis anymore—it’s just a receipt. This shift unlocks massive change: • Imaging gets faster, cheaper, safer (goodbye excess radiation) • Screening moves upstream—AI finds risk in heartbeat patterns, retinal scans, voice recordings • Diagnosis becomes multi-modal, AI-native, and decentralized So no—AI isn’t replacing radiologists. It’s replacing the idea of radiology.

One of the most exciting shifts happening in radiology right now is that AI is no longer sitting on the sidelines, it’s being embedded directly into the reporting workflow. Solutions like PowerScribe One and Microsoft Dragon Copilot are designed with that exact principle in mind—bringing AI into the natural flow of how radiologists work. Instead of adding another tool to manage, AI is integrated to assist in real time: -Automating documentation and reducing administrative burden -Enhancing report accuracy and consistency -Accelerating turnaround times so clinicians can act faster -Allowing radiologists to stay focused on interpretation and patient care What makes this even more impactful is how these capabilities are being shaped—not in isolation, but through deep collaboration with healthcare organizations and partners. Real-world deployment, feedback, and iteration are driving meaningful, practical innovation that fits seamlessly into clinical environments. This is where AI moves from concept to measurable clinical value.