Artificial Intelligence Is Redefining Radiology from the Inside Out

The Radiologist’s New Eyes

In 1895, Wilhelm Röntgen discovered that a cathode ray tube could make the bones in his wife’s hand visible through flesh. It took him seven weeks to produce the first X-ray image. Bertha Röntgen had to hold her hand still for fifteen minutes. Today, a CT scanner can capture thousands of slices of a human body in under a minute, and a radiologist might interpret hundreds of such scans in a single shift. But here is the paradox nobody tells you: the harder we look, the more we miss.

That is the problem that Reabal Najjar, a researcher at the University of Jordan, lays out in a 2023 review published in Diagnostics. Najjar argues that radiology has reached a threshold where human perception alone cannot keep pace with the volume and complexity of medical imaging (Najjar, 2023). The field generates exabytes of data every year. A single CT scan can contain more than 3,000 images. And the human eye, for all its evolutionary brilliance, was not designed to spot a three-millimeter lung nodule hidden in a sea of gray pixels at 2 AM on a Tuesday.

Enter artificial intelligence. Not as a replacement for radiologists, but as something stranger. Something that is changing the specialty from the inside.

What Actually Happens When a Machine Looks at a Scan

Najjar’s review is not a single experiment. It is a synthesis of hundreds of studies published between 2015 and 2023, covering the entire arc of AI’s infiltration into radiology. The author traces how machine learning and deep learning models have moved from academic curiosities to tools that now perform image segmentation, computer-aided diagnosis, predictive analytics, and workflow optimization in real clinical settings (Najjar, 2023).

Let me translate that.

Image segmentation means the AI draws boundaries around structures. A tumor. An organ. A blood vessel. What once took a radiologist minutes of manual tracing now happens in seconds. Computer-aided diagnosis means the AI flags suspicious regions. Not a diagnosis, but a suggestion. A second pair of eyes that never gets tired, never gets distracted, never skips lunch.

The review draws on case studies across multiple medical disciplines. In chest radiology, deep learning models have matched or exceeded radiologist performance in detecting lung nodules on CT scans. In mammography, AI systems have reduced false positives by up to 30 percent in some studies. In neuroradiology, convolutional neural networks can identify intracranial hemorrhages faster than the average resident (Najjar, 2023).

But here is the finding that stopped me: Najjar reports that the most dramatic gains come not from replacing humans, but from augmenting them. When radiologists worked with AI assistance, their diagnostic accuracy improved more than when either worked alone. The whole was genuinely greater than the sum of its parts.

The Black Box Problem Nobody Wants to Talk About

Here is where the story gets uncomfortable.

Najjar spends a significant portion of the review on what the author calls the “black box” problem (Najjar, 2023). These AI models, particularly deep learning networks, make decisions in ways that are often opaque even to the engineers who build them. A neural network might correctly identify a malignant tumor, but it cannot always explain why it flagged that particular patch of pixels. It learned patterns from thousands of scans, but those patterns are encoded in millions of weighted connections that no human can trace.

This matters because radiology is not just about being right. It is about being accountable. When a radiologist misses a finding, there is a person to question, a reasoning process to examine, a training history to review. When an AI misses a finding, or worse, when it makes a mistake that leads to a misdiagnosis, who is responsible? The algorithm? The developer? The hospital that deployed it?

Najjar does not answer this question. The author frames it as an open ethical and legal challenge that the field has not yet resolved (Najjar, 2023). And this is not a theoretical concern. In 2021, a study found that commercial AI systems for chest X-ray interpretation showed significant performance drops when tested on data from hospitals different from their training sources. The models were brittle. They learned shortcuts, not anatomy.

The black box problem is not just philosophical. It is practical. If a radiologist cannot trust an AI’s reasoning, they cannot rely on it. And if they cannot rely on it, they will not use it. The technology stalls.

Why Your Doctor’s Workflow Is About to Change

Najjar’s review devotes a substantial section to workflow optimization, and this is where the author’s analysis feels most urgent (Najjar, 2023). Radiology departments are bottlenecked. The number of imaging studies performed each year grows by roughly 5 to 10 percent, but the number of radiologists grows by maybe 1 percent. Burnout rates among radiologists are among the highest in medicine.

AI can help in ways that have nothing to do with diagnosis.

Triage algorithms can scan incoming studies and flag the most urgent cases. A patient with a suspected stroke gets priority over a patient with a chronic back problem. The AI does not make the diagnosis. It just says: look at this one first. One study cited in the review found that AI-powered triage reduced turnaround time for critical findings by nearly 40 percent (Najjar, 2023).

Then there is the tedium. Radiologists spend a significant portion of their day doing repetitive tasks. Measuring nodules. Tracking changes over time. Comparing current scans to previous ones. These are tasks that AI can automate with high reliability, freeing radiologists to focus on the complex cases that require human judgment.

But Najjar warns that automation comes with its own risks. If radiologists become too reliant on AI, they may lose the skills that make them good radiologists. Pattern recognition is a muscle. If you stop using it, it atrophies. The author calls for careful integration, where AI handles the routine but radiologists remain actively engaged in the interpretive process (Najjar, 2023).

The Data Quality Trap

Here is a detail from the review that surprised me. Najjar writes that one of the biggest obstacles to AI integration is not technical capability but data quality (Najjar, 2023). Medical imaging data is messy. Different hospitals use different scanners. Different technicians use different protocols. A scan done at a large academic center might look completely different from a scan done at a rural clinic, even for the same condition.

AI models trained on pristine, high-quality data from elite institutions often fail when deployed in the real world. The images are noisier. The patient populations are different. The equipment is older. This is called distribution shift, and it is one of the reasons why many promising AI tools never make it past the research phase.

Najjar’s review cites studies showing that performance drops of 10 to 20 percent are common when AI models are tested on data from sites not included in their training (Najjar, 2023). This is not a small problem. It is a fundamental limitation that the field is only beginning to address through techniques like domain adaptation and federated learning.

What the Research Does Not Prove

It would be easy to read Najjar’s review and conclude that AI is about to take over radiology. That is not what the evidence shows.

The review does not prove that AI can replace radiologists. It does not even suggest that this is the goal. Every case study included in the analysis shows AI performing best when paired with human expertise. The machines are good at pattern recognition. They are bad at context. They cannot talk to a patient. They cannot consider a family history. They cannot make a judgment call when the imaging findings conflict with the clinical picture.

The review also does not prove that AI systems are safe for widespread deployment without rigorous oversight. Najjar is careful to note that most of the studies cited are retrospective. They test AI on data that has already been interpreted by humans. Prospective trials, where AI is used in real time on real patients, are far rarer (Najjar, 2023). The gap between research and clinical reality remains wide.

And the review does not address one question that keeps me up at night: what happens when an AI system is trained on biased data? If the training data comes predominantly from one demographic group, the model will perform worse on others. This is not a hypothetical. It has happened with dermatology AI, with pulse oximeters, with kidney function algorithms. Radiology AI is not immune.

The Collaboration That Could Change Everything

Najjar’s most provocative argument comes at the end of the review. The author advocates for a new kind of partnership between radiologists and AI developers, one that goes beyond the usual consultant relationship (Najjar, 2023). Radiologists should not just be users of AI tools. They should be co-creators.

This means radiologists learning enough about machine learning to understand what the models are doing and what their limitations are. It means AI developers spending time in reading rooms, watching how radiologists actually work. It means building tools that fit into existing workflows rather than demanding that workflows be redesigned around the tools.

The review cites examples where this kind of collaboration has worked. At one institution, radiologists worked with engineers to develop a model for detecting pancreatic cancer on CT scans. The radiologists provided the clinical expertise. The engineers provided the technical skills. The result was a system that caught tumors that human readers had missed, without generating excessive false positives (Najjar, 2023).

This is the model that Najjar wants to scale. Not AI as a black box dropped into a department, but AI as a tool shaped by the people who will use it.

What This Actually Means

• ▸Radiologists should start learning basic machine learning concepts now. The field is moving toward a model where radiologists are expected to understand what their AI tools are doing, not just accept their outputs. This is not optional. It is becoming a core competency.

• ▸Hospitals should test AI systems on their own data before deploying them. The performance numbers from published studies are not reliable predictors of how a system will work in a specific clinical setting. Local validation is not a luxury. It is a necessity.

• ▸AI is best used for triage and repetitive tasks, not for final diagnosis. The evidence consistently shows that human AI collaboration outperforms either alone. The goal should be to free radiologists for complex cases, not to remove them from the loop.

• ▸The black box problem is not going away. Until AI systems can explain their reasoning in terms that humans can verify, they should be treated as assistants, not authorities. Radiologists must maintain the final interpretive responsibility.

• ▸Data quality is the bottleneck. The biggest gains in AI performance will come not from better algorithms but from better data. Standardized imaging protocols, shared datasets, and rigorous quality control are the hidden infrastructure that will make AI work.

The X-ray changed medicine because it let doctors see what they could not see before. AI is doing the same thing, but this time the invisible thing is not bone. It is pattern. It is probability. It is the signal buried in noise that no human eye can parse. The difference is that this time, the tool also sees us. It watches how we work, learns from our decisions, and changes itself in response. That is not just a new technology. It is a new kind of relationship between doctor and machine. And we are only beginning to understand what that relationship will demand.