Schell Introduction to Medical Open Network for AI

Introduction to MONAI, the Framework Hospitals, Radiologists, use to Turn Medical Images Into Clinical Intelligence
1. Auflage 2026
ISBN: 979-8-90498-465-6
Verlag: PublishDrive
Format: EPUB
Kopierschutz: 0 - No protection

E-Book, Englisch, 266 Seiten

Medical imaging AI is real work now - segmenting tumors on MRI, catching strokes on CT, measuring cardiac function on echo - and most of it, quietly, runs on one open-source framework: MONAI. This book takes you inside MONAI: what it is, how it works, and how a model actually goes from a Jupyter notebook to a hospital PACS system.
Follow Cerulean Imaging AI - a fictional composite of the data engineers, radiologists, MLOps, and quality specialists behind a modern medical imaging AI product - through every real step.
Inside you will learn:
• Why medical imaging AI, and why now - the scanner problem, the four tasks, the model-to-practice gap
• What MONAI actually is - core library, MONAI Label, MONAI Deploy, MAPs, Model Zoo, Auto3DSeg
• How a medical imaging model works - voxels, DICOM, convolutions, the U-Net, the training loop
• A short history - the classical era, the 2012 shift, U-Net in 2015, fragmented codebases, MONAI's arrival
• The data problem - scanner to PACS to model, labeling cost, bias, privacy, provenance, human-in-the-loop
• Building a model with MONAI - naming the task, splitting scans not slices, transforms, architecture choice
• Deploying to a hospital - notebook-to-hospital gap, packaging as a MAP, on-prem vs cloud, sitting next to PACS
• Validation, regulation, and safety - FDA pathway, monitoring, drift
• Where MONAI is going - the open consortium, foundation models, multi-modal imaging
• The credible next decade - where the field is honestly headed
Ten chapters. Framework first, hype-free. The MONAI field guide for anyone building medical imaging AI in 2026 and beyond.

Autoren/Hrsg.

Schell, Victor

Leseproben

Why Medical Imaging AI, and Why Now

It is 2:14 on a Tuesday morning in a windowless reading room three floors under a hospital in the suburbs. Dr. Maya Okonkwo, a neuroradiologist by training and the co-founder and CEO of a five-person startup called Cerulean Imaging AI, is on call. A stroke alert flashes in her queue. The scan is a non-contrast head CT on a 62-year-old man who arrived at the emergency department twenty minutes earlier, awake but confused, right arm heavy, speech thickening at the edges. Maya has read thousands of these. She knows the shape of what she is looking for, and she knows the shape of what she is likely to miss when she is tired.

She scrolls. The obvious findings are not there. No hyperdense vessel sign, no clear loss of gray-white differentiation, no big edema. She is ready to call the study negative and phone the emergency physician when a small blue outline blinks on her monitor. A software tool Cerulean built on top of the open-source MONAI framework flags a subtle asymmetry in the left basal ganglia. Maya widens the window, changes the contrast, and looks again. The asymmetry is real. It is the earliest imaging sign of a middle cerebral artery infarct, and at this hour, with this exact appearance, on this exact patient, she is not sure she would have seen it on her own.

She calls the stroke team. Within eleven minutes, the patient is on the CT angiography table, and within forty minutes, an interventional neuroradiologist is threading a catheter to open the blocked artery. Maya walks back to her office, sits down, and does something she has done a lot since starting Cerulean. She writes a two-line note to her co-founder Rohan Bhatt, the machine learning lead. The note says, roughly, that the tool worked, that the case would probably have had a good outcome either way, and that she still isn't sure whether she should be grateful, humbled, or a little bit terrified. Probably all three.

This is a book about what makes moments like that possible, about the framework, called MONAI, that a lot of teams like Cerulean now build on, and about the honest limits of that framework. Before any of that, you need the map. You need to know why medical imaging AI became a real field instead of a set of conference posters, why it took as long as it did, what it can actually do today, and where the field is quietly stuck. That map is what this opening chapter gives you.

Cerulean itself is fictional. The company, its people, and every clinical scenario in this book are illustrations, not case studies. Nothing here is guidance for building or deploying real clinical software, and you should treat every example as a story that helps you see the shape of the field, not as a template. With that understood, here is the ground you are standing on.

The Scanner Problem That Will Not Go Away

Start with a boring fact that turns out to be the reason this whole field exists. Medical scanners keep getting better. In this context, better means they produce more data. A chest X-ray from 1985 is a single image on film. A chest CT from 2005 might be 150 slices. A modern multi-detector cardiac CT can be a thousand slices per phase, across several phases, in a single study. If you print that study, you get a tall stack of paper. If you read it on a workstation, you are scrolling for a long time.

Diagram 1.1.

Your visual system did not evolve for this. A trained radiologist is genuinely remarkable at spotting the abnormal inside the normal, but attention is a finite resource. When the number of images per study goes up by a factor of ten, the time a radiologist has to spend on each study does not go up by a factor of ten. In many practices, it barely moves. The reading list stays the same length. The workday stays the same length. The scans get bigger, denser, and more numerous. Something has to give.

For a long time, what gave was implicit. Radiologists compensated with pattern recognition, search strategies, and hard-won experience. They learned which slices to check first, which windows to use, which subtle findings to look for in which clinical contexts. That skill is real, and it is not going anywhere. But it is also stretched thin, and not because anyone is doing anything wrong. The input has grown faster than the human system can keep up with.

Consider what a single hospital shift can look like today. An overnight radiologist covering a busy trauma center may read two hundred studies. That includes head CTs, cervical spine CTs, chest CTs for pulmonary embolism, abdominal CTs for appendicitis and diverticulitis, chest X-rays after line placements, and a handful of MRIs. Each study is a moving stack of images. Each is a decision that will change what happens to a patient in the next hour. The volume is not a bug of the modern hospital. It is a feature of good imaging being cheap enough that clinicians order it whenever they think it might help.

That pressure made medical imaging AI more than an academic exercise. Software does not get tired at 4 a.m. It does not blink. It can scan every slice of every study at the same level of attention as the first one, and it can do that in seconds. That doesn't mean software replaces the reader. It means the reader can be given a first pass, a second opinion, a triage signal, or a measurement that would have cost precious minutes to compute by hand. Multiply small time savings by a real hospital's daily volume, and the number grows fast.

A subtler version of the same pressure exists in subspecialty imaging. In neuro-oncology, tracking a brain tumor across time means measuring its volume on each new MRI. Doing that by hand, slice by slice, takes a radiologist twenty to forty minutes per study if it is done seriously. That's why, in practice, tumor sizes are often approximated by measuring the longest diameter on one slice. That approximation is defensible for many uses, but it is not the same as knowing the volume. Software can produce a real volume in under a minute, every time, on every scan.

You can walk the same story through mammography, lung screening, cardiac imaging, prostate MRI, and dental cone-beam CT. In every one of those areas, the volume of data per patient has grown, the number of trained specialists has not grown at the same rate, and pressure to get more consistent measurements from each study has intensified. That is the demand side of medical imaging AI. It is not a hypothesis about the future. It is the day-to-day economics of every large radiology practice on the planet.

I open this book with the scanner problem, rather than a description of neural networks, because the scanner problem is why the field isn't going away, even when the hype does. If every AI startup collapsed tomorrow, the data problem would still be here. Someone would still have to figure out how to help human readers cope with studies that are ten and a hundred times larger than they used to be. Modern machine learning became the first tool with a real chance of doing that at scale. That seed became the start of everything that followed.

What Changed Between 2012 and Today

Medical image analysis is not new. Researchers have been writing programs to segment brain scans, count cells, and measure vessel diameters since at least the 1980s. If the problem is old, and the demand is old, then something specific must have changed to turn a slow academic corner into a fast commercial one. That something is not a single breakthrough. It is three separate curves that crossed at almost the same time.

Diagram 1.2.

The first curve is a family of methods called deep learning. For decades, computer vision worked by hand-crafting features. A researcher would decide that edges, corners, and specific texture patterns were the important signals, and then a machine learning model would decide how to combine them. That worked for controlled problems and struggled in the wild. Deep learning turned the problem on its head. Instead of telling the model what features to look at, you show it many labeled examples and let it learn its own features, at many levels of abstraction, from raw pixels up. The shift from hand-crafted to learned features is why a system can now reliably tell a cat from a dog, and why a similar system can tell a stroke from normal brain tissue with real accuracy.

The public turning point was 2012. A neural network called AlexNet won a large image classification contest by a wide margin, using a design that would have looked strange three years earlier. That result kicked off a decade of rapid improvement, first in general images, then in specialized domains like medical imaging. The methods didn't appear out of nowhere. They rested on ideas quietly developed since the 1980s. But in 2012, everyone in the room noticed at once that something had shifted.

The second curve is hardware. Deep learning is arithmetic-heavy, and the arithmetic is highly parallel. It turned out that the graphics processing units, or GPUs, built for video games were almost perfectly suited to that arithmetic. A single modern GPU can perform trillions of operations per second on the kind of matrix math these models require. That kind of power was not affordable to a graduate student in 2005. By the mid-2010s, it was. By the early 2020s, cloud providers were renting time on eight-GPU nodes by the hour, and even a small startup could train a serious model over a weekend for the cost of a modest dinner.

The third curve is data, especially open data. Medical images are difficult to collect and share because they are private, regulated, and expensive to label. Even so, over the last fifteen years, a series of open datasets appeared that were large enough to matter. Public chest X-ray collections with tens of thousands...