From Pixels to Predictions: How AI Found Its Place in the Radiology Suite

By Professor & Dr. Anthony Noujaim Endowed Chair of Oncology. Director, Division of Oncologic Imaging and Radionuclide Therapy, Faculty of Medicine & Dentistry, University of Alberta

A brief note on terminology: where specialist shorthand is unavoidable, terms are defined at first use. A compact glossary appears at the end of this article.

The story of artificial intelligence in radiology did not begin with the flashy deep-learning breakthroughs of 2012 that dominate most popular retellings. It started quietly — in the late 1990s, in overlit reading rooms where radiologist headcounts were flat while imaging volumes were doubling. The field did not wait for Silicon Valley to arrive; it built its own first generation of machine readers. And it did so because radiology, almost uniquely in medicine, already had everything an intelligent algorithm needs: standardized digital data, pattern-based diagnosis, and measurable outcomes.

That thirty-year journey — from brittle rule-based detectors to today’s generalist AI platforms — is the foundation every healthcare executive must understand before committing capital to the next generation of imaging AI.

Even as interpretation tools advanced, AI was quietly infiltrating every other stage of the imaging workflow.

Why Radiology Was Always AI’s Natural Home

Three structural characteristics made radiology medicine the earliest and most receptive AI frontier.

Data abundance. By the mid 1990s, Picture Archiving and Communication Systems (PACS — the digital libraries where scans are stored and retrieved) had converted X-rays, CT (computed tomography) scans, and mammograms into a universal machine-readable file format called DICOM. Unlike surgical video or tissue slides, radiology produces millions of consistently formatted digital images every second: precisely the structured, labeled, comparable data that machine learning requires at scale.

Pattern primacy. Radiological diagnosis is, at its core, a pattern-recognition task — identifying nodule margins, lesion density, bone displacement, deviations from normal anatomy. Human eyes are excellent pattern recognizers, but as caseloads mount and shifts lengthen, fatigue gradually erodes vigilance — and it is that accumulated fatigue, not inherent capability, that creates the conditions under which subtle findings can go undetected. The U.S. National Lung Screening Trial (NLST) put a precise number on what radiologists already knew intuitively: even in a rigorously controlled research setting, a meaningful proportion of small but actionable lung nodules — in the 4–6 mm range — were not flagged on first review. [1]

Measurability. This is where radiology has always had an edge that other specialties quietly envied. Sensitivity and specificity were never novel concepts in our field — they were the native language of the reading room, the standard by which every technique, every protocol, every piece of equipment was judged, long before AI entered the conversation. A surgeon’s skill, an internist’s clinical judgment, a psychiatrist’s diagnostic intuition: none of these reduce neatly to a performance curve. Radiology did — and so, to their credit, did laboratory medicine. But outside those two disciplines, medicine largely evaluated itself through consensus, experience, and outcome proxies. That quantitative culture turned out to be exactly what regulators needed. The FDA did not have to invent a new framework to evaluate AI imaging tools; the evidentiary standard was already there, embedded in decades of diagnostic performance literature. For a technology looking for a home in medicine, that was no small thing — it meant a credible path from algorithm to approved product, at a time when most of healthcare was still arguing about what ‘evidence’ even meant for software.

The CAD Era: When Algorithms Were Rule-Books (Late 1990s–2010)

In 1998, R2 Technology received FDA clearance for the ImageChecker M1000 — the world’s first commercial computer-aided detection (CAD) system, designed to flag suspicious areas on mammograms as a second opinion for radiologists. It was not deep learning; it was hand-coded pattern matching, running on single-processor workstations comparable in power to a high-end desktop PC of the era. [2]

Siemens Healthineers advanced the concept with syngo.CT Lung CAD (FDA-cleared 2006, updated 2015), a tool designed to automatically mark potential solid pulmonary nodules on chest CT scans — including nodules as small as 3 mm and those adjacent to blood vessels or the chest wall, where human detection is hardest. In multi-center studies, it demonstrably increased radiologists’ detection accuracy for clinically significant nodules. [3]

Performance was honest rather than transformative. Sensitivity hovered around 70–75%; false-positive rates were high enough that radiologists spent meaningful time dismissing phantom findings. The NLST data showed 10–20% detection gains for very small nodules — useful, but not the step-change the field needed. These tools demonstrated the fundamental idea of AI as enhancement rather than replacement and made it abundantly evident that the limiting factor was processing power rather than intellectual ambition.

The Deep Learning Inflection: Graphics Cards Rewrite the Rules (2010s)

The pivotal moment arrived not from a hospital but from an image-recognition competition. In 2012, a deep neural network called AlexNet, developed at the University of Toronto, reduced the leading error rate in the annual ImageNet visual classification challenge by more than 10 percentage points in a single year — an improvement that had taken the previous five years to accumulate collectively. It did so by running on NVIDIA graphics processing units (GPUs), chips originally designed for video games but ideal for the massively parallel calculations that deep neural networks require. Training that had taken weeks on a standard computer could now run in days. [4]

Radiology adopted the architecture within two years. Deep convolutional neural networks (CNNs, multi-layered image-analysis architectures that learn features automatically from pixels) consumed large public datasets such as ChestX-ray14 (112,000 chest X-rays) and began matching radiologist-level performance on specific tasks: detecting intracranial hemorrhage, grading diabetic eye disease, and identifying lung nodules. By 2015, multi-GPU clusters allowed models to train on millions of imaging slices, a scale unimaginable a decade earlier.

The limitations were equally instructive. Each model needed tens of thousands of expert-labeled images per task, per anatomy, per scanner manufacturer. A hemorrhage-detection model trained on scans from one hospital system frequently underperformed at another with different equipment or patient demographics. The brittleness that had constrained 1990s CAD tools had not disappeared, it had simply scaled up.

Computing Power: The Silent Co-Author of Every Advance

Radiology AI’s evolution cannot be separated from the hardware that made it possible:

Era	Hardware	Speed	AI Milestone
Late 1990s	Single-CPU workstations	~100 MFLOPS¹	Rule-based CAD; ImageChecker
2006–2012	Multi-core CPUs + early GPUs	~1 TFLOPS²	syngo Lung CAD; first CNN experiments
2012–2018	NVIDIA Kepler/Pascal GPUs	~10 TFLOPS	AlexNet, ResNet, U-Net in clinical trials
2020–present	A100/H100 GPU clusters	~1 PFLOPS³	Foundation models trained on millions of scans

¹ MFLOPS: millions, ² TFLOPS: trillions, ³ PFLOPS: quadrillions of mathematical operations per second — measures of raw computing speed.

Each generational leap compressed model-training time from weeks to hours, enabling progressively more sophisticated networks. Crucially, the universal DICOM standard for radiology allowed information from many hospitals and scanner brands to be combined and shared, a structural advantage that other clinical disciplines could not match due to proprietary and incompatible data formats.

AI Expands Beyond the Reading Room

Even as interpretation tools advanced, AI was quietly infiltrating every other stage of the imaging workflow.

At the front end, clinical decision-support systems embedded the American College of Radiology’s (ACR) Appropriateness Criteria — evidence-based guidelines for which scan is right for which clinical question — directly into physicians’ ordering screens, measurably reducing inappropriate imaging requests. Automated scan protocol selection, using machine learning to translate a referring physician’s clinical question into the correct imaging sequences, achieved greater than 95% accuracy in early deployments, cutting the radiologist’s time spent on administrative protocolling. [5]

At the back end, deep learning-based image reconstruction transformed what was physically achievable at the scanner itself. GE Healthcare’s TrueFidelity and Canon Medical’s AiCE (Advanced intelligent Clear-IQ Engine), both FDA-cleared by 2020, used neural networks trained on high-quality reference scans to separate true anatomical signal from image noise. Clinical studies showed radiation dose reductions of 38–71% while maintaining or improving diagnostic image quality.[6]

Three Decades of Lessons Every CXO Should Keep

The thirty-year arc from ImageChecker to today’s foundation models yields four principles that have proven durable across every generation of technology:

1. Augmentation outperforms automation. The tools that gained adoption helped radiologists work better and faster. Those that attempted to operate independently stalled in regulatory review and clinical resistance.
2. Infrastructure precedes innovation. GPU availability — not algorithmic genius — determined which ideas became clinical products. Organizations that invested in computing infrastructure first moved fastest.
3. Workflow integration beats benchmark accuracy. A tool that achieves 95% accuracy but disrupts a radiologist’s reading flow will be abandoned. A tool with slightly lower accuracy that fits seamlessly into the existing screen and process will be used millions of times a year.
4. Data diversity beats data volume. The 1990s CAD failures on underrepresented edge cases are echoed today in AI tools that underperform on specific patient populations, scanner types, or clinical presentations. Breadth of training data matters more than raw quantity.

These hard-won lessons are precisely what motivated the next generation: foundation models — large, generalist AI platforms pre-trained on diverse, multi-modal datasets and adaptable across anatomies, imaging modalities, and clinical tasks without being rebuilt from scratch for each one. That transformation, and what it means for how radiology departments — and their technology partners — operate today, is the subject of Part 2.

Quick Glossary

Term	Meaning
PACS	Picture Archiving and Communication System — the digital library where scans are stored and retrieved
DICOM	Digital Imaging and Communications in Medicine — the universal file format for medical images
CAD	Computer-Aided Detection/Diagnosis — algorithms that flag findings for radiologist review
CT	Computed Tomography — cross-sectional X-ray imaging
CNN	Convolutional Neural Network — a deep learning architecture optimized for image analysis
GPU	Graphics Processing Unit — the parallel-computing chip powering modern AI training
NLST	National Lung Screening Trial — landmark U.S. study on low-dose CT lung cancer screening
ACR	American College of Radiology
Foundation Model	A large AI pre-trained on vast datasets, adaptable to many tasks with minimal retraining

---

References

1. National Lung Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med. 2011;365(5):395–409.
2. Nakahara H, Namba K, Fukami A, Watanabe R, Mizutani M, Matsu T, Nishimura S, Jinnouchi S, Nagamachi S, Ohnishi T, Futami S, Flores II LG, Nakahara M, Tamura S. Computer-Aided Diagnosis (CAD) for Mammography: Preliminary Results. Breast Cancer. 1998 Oct 25;5(4):401-405. doi: 10.1007/BF02967438. PMID: 11091682.
3. Siemens Healthineers. syngo.CT Lung CAD — clinical applications [product information]. Erlangen: Siemens Healthineers; 2015. Available from: siemens-healthineers.com.
4. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Adv Neural Inf Process Syst. 2012;25:1097–1105.
5. Giess CS, Ip IK, Schneider L, Hanson R, Imsirovic H, Khorasani R. Influence of patient-centered clinical decision support on appropriate imaging utilization. J Am Coll Radiol. 2014;11(7):677–683.
6. Greffier J, Dabli D, Hamard A, Belaouni A, Akessoul P, Frandon J, Beregi JP. Effect of a new deep learning image reconstruction algorithm for abdominal computed tomography imaging on image quality and dose reduction compared with two iterative reconstruction algorithms: a phantom study. Quant Imaging Med Surg. 2022 Jan;12(1):229-243. doi: 10.21037/qims-21-215. PMID: 34993074; PMCID: PMC8666764.

Next in this series — Part 2: “The Generalist Machine: How Foundation Models Are Reshaping the Entire Imaging Chain.”